Lithium ion battery capacity prediction method based on random charging fragments and indirect characteristics

Abstract

The invention relates to a lithium ion battery capacity prediction method based on random charging fragments and indirect characteristics, and belongs to the technical field of power batteries. Feature analysis and indirect feature acquisition; feature 2 prediction based on an extreme random forest algorithm; feature 2 prediction result analysis; predicting the capacity of a single battery; and carrying out model generalization verification. Charging time in two specific voltage ranges in the constant current charging process is used as a health characteristic HF1 and a health characteristic HF2. The whale optimization algorithm is enhanced to be applied to hyper-parameter optimization of Gaussian process regression, a single battery capacity prediction model is constructed, and the number of training sets is greatly reduced. According to the method, the characteristic factors with high correlation can be directly extracted from the charging process, the original measurement data are directly input into the established capacity prediction model without being influenced by noise, the capacity of different types of batteries can be accurately predicted under the condition that the data volume is limited, and the generalization is good.

Description

Technical Field

[0001] The present invention belongs to the technical field of power batteries, and in particular relates to a lithium-ion battery capacity prediction method based on random charging segments and indirect features. Background Art

[0002] As a clean and efficient energy storage solution, lithium-ion batteries have been promoted in electric vehicles. However, lithium batteries will inevitably experience capacity decay during use, which directly affects the reliability of the battery. Therefore, accurate estimation of battery capacity is of great significance to improving the safety and performance of electric vehicles. The existing technology is basically based on a data-driven method, which does not require complex physical modeling or research on the internal aging mechanism of the battery. It only needs to use the data obtained during the battery charging and discharging experiment stage to achieve accurate capacity prediction.

[0003] However, traditional data-driven methods not only have high requirements on feature parameters, but also require a large amount of training data. In addition, when the battery feature data is missing during the collection process, the established model may not be usable, and the battery data needs to be re-analyzed and features need to be selected. Therefore, how to reasonably obtain aging features based on short-term charging data and random charging fragments, and improve the prediction accuracy of power battery capacity while reducing the amount of training data, and achieve accurate prediction of different battery capacities, is a key problem that needs to be solved. Summary of the invention

[0004] To solve the above problems, the present invention provides a lithium-ion battery capacity prediction method based on random charging segments and indirect features.

[0005] In order to solve the above technical problems, the technical solution of the present invention is to provide a lithium-ion battery capacity prediction method based on random charging segments and indirect features, which specifically includes the following steps:

[0006] S1. Perform cyclic charge and discharge tests on pre-set lithium-ion batteries of different models until the discharge capacity of each model of battery is less than 80% of the nominal capacity, and record the charge and discharge data in real time;

[0007] The pre-set different types of lithium-ion batteries are: different batteries of the same type and different batteries of different types;

[0008] The charging and discharging data include: charging voltage, discharging capacity and charging time;

[0009] S2. Obtain two health characteristics of the lithium-ion battery through two preset voltage intervals, and obtain a data set of battery health characteristics and battery capacity according to Spearman correlation analysis;

[0010] The two preset voltage ranges are: 3.55V~3.75V and 3.80V~4.15V;

[0011] The two health features of lithium-ion batteries are health feature 1 (HF1) and health feature 2 (HF2). Health feature 1 (HF1) is regarded as an indirect feature to predict health feature 2 (HF2). Spearman correlation analysis is used to verify the correlation and obtain the health feature and capacity data set.

[0012] S3. Predict the health feature 2 (HF2) based on the extreme random forest algorithm to obtain the predicted feature (HF2'), the steps are as follows:

[0013] S3.1, HF1 and HF2 are subjected to Cox-Box transformation to enhance the linear correlation between HF1 and HF2, and HF10 and HF20 are obtained after the transformation; the calculation formula of Cox-Box transformation is as follows:

[0014]

[0015] In the formula, y i,k For each data point, we need y i,k >0,k=1,2,corresponding to HF1 and HF2;In(y i,k ) is for y i,k Take the logarithm; λ is a constant, and the maximum likelihood estimation method is needed to determine the λ value suitable for the data;

[0016] S3.2, randomly select HF10 and HF20 corresponding to 30% of the Cox-Box transformations, use HF10 as input and HF20 as output, and input them into the extreme random forest model for training; then input HF10 corresponding to the remaining 70% of the cycles into the trained extreme random forest model to obtain the predicted value HF20';

[0017] S3.3, perform Cox-Box inverse transformation on HF20' to obtain the prediction feature (HF2');

[0018] The calculation formula for the inverse Cox-Box transformation is as follows:

[0019]

[0020] In the formula, For Get the index;

[0021] S4, relative error analysis and Spearman correlation coefficient analysis were performed between the predicted feature (HF2') and the healthy feature 2 (HF2) to verify the prediction accuracy;

[0022] The expression of Spearman correlation coefficient analysis is as follows:

[0023]

[0024] In the formula, i=1, 2 represent HF1 and HF2 respectively; n represents the total number of samples, that is, the total number of HF or SOH; k represents the number of values; x represents HF; y represents the capacity; and Corresponding to the means of x and y respectively;

[0025] S5, predicting the capacity of a single battery;

[0026] Specifically, the indirect features and predicted features are used as input, and the battery capacity is used as output, and are input into a gated recurrent unit neural network model based on an enhanced whale optimization algorithm for training. The training set is determined to be the data of the first 20% of the cycles, and the validation set is determined to be the data of the remaining 80% of the cycles, and the optimal capacity prediction model is saved;

[0027] The prediction steps are as follows:

[0028] S5.1. Set the network structure of the gated recurrent unit neural network and initialize the neural network hyperparameters;

[0029] S5.2, select any battery data in HF1 and the battery data corresponding to HF2' as the input of the gated recurrent unit neural network, use the capacity of the battery collected by S1 as the output of the neural network, determine that the training data is HF1, HF2' and capacity corresponding to the first 20% of the cycles, and the verification data is the data of the remaining 80% of the cycles;

[0030] S5.3, taking the mean square error of the capacity prediction value as the objective function, and defining the optimization objective as minimizing the mean square error of the capacity prediction value;

[0031] Specifically, the optimization object is set as the hyperparameters of the gated recurrent unit neural network, including: the number of iterations, the learning rate, and the number of neurons in the hidden layer; the enhanced whale optimization algorithm parameters are set, including: the number of whales, the maximum number of iterations, and the upper and lower limits of the optimization object;

[0032] S5.4. Evaluate the performance of the model based on the enhanced whale optimization algorithm;

[0033] Specifically, a group of whales (i.e., candidate solutions of the algorithm) are randomly initialized in the hyperparameter space; a gated recurrent unit neural network model is trained with the hyperparameter set represented by each whale, and the performance of the model is evaluated using the validation set;

[0034] The position of the whale (i.e., the optimal hyperparameter) is updated according to the rules of the enhanced whale optimization algorithm. The algorithm will simulate the predation behavior of whales, mainly including three stages: surrounding prey, spiral attack, and searching for prey, and optimize individuals to approach the optimal solution. Compared with the classic whale optimization algorithm, the enhanced whale optimization algorithm improves the performance of the classic whale optimization algorithm by introducing a pooling mechanism and three effective search strategies (migration search strategy, priority search strategy, and surrounding prey search strategy).

[0035] The pooling mechanism is defined as: given a matrix of size k, Pool = (p1, p2, ..., p k ), at the end of each iteration, the element P in the matrix Pool is generated m =(p m,1 ,p m,2 ,...,p m,D ), wherein, in this embodiment, k=5;

[0036] Element P m The resulting formula is as follows:

[0037]

[0038] In the formula, Represents the element of the tth iteration; is the binary random vector in the tth iteration, is its inverse vector; is the worst solution in the tth iteration; It represents the optimal solution generated in the tth iteration A random position adjacent to the corresponding whale; where t=20 in this embodiment; at the end of each iteration, the pooling mechanism crosses the worst solution in the current iteration with a randomly generated promising solution to increase population diversity; this mechanism improves the algorithm's search capability by mixing the worst solution with the promising solution, avoiding premature convergence and local optimal traps;

[0039] The expression of the migration search strategy is as follows:

[0040]

[0041] Where: represents the position of the hth whale in the t+1th iteration; represents the random position of the whale in the search space in the tth iteration; rand is a random number between [0,1) and it follows a uniform distribution; δ max and δ min Corresponding to the upper and lower bounds of the optimization object respectively; δ best-max and δ best-minThen they correspond to the optimal solutions The migration search strategy increases exploration ability and population diversity by randomly separating a part of whales to cover unvisited areas. This strategy improves the algorithm's ability to escape local optimality by introducing randomness.

[0042] The expressions for the preferred search strategy are as follows:

[0043]

[0044] Where: represents the position of the hth whale in the tth iteration; and are positions 1 and 2 randomly selected from the matrix Pool in the tth iteration; MaxIt represents the maximum number of iterations set in the enhanced whale optimization algorithm; in order to improve the exploration ability of the classic whale optimization algorithm in searching for prey, the enhanced whale optimization algorithm adopts a priority selection search strategy, which discovers diverse solutions by dispersing whales in different areas of the search space, so a larger step size is required;

[0045] The expression of the encirclement prey search strategy is as follows:

[0046]

[0047] Where: is the position 3 randomly selected from the matrix Pool in the tth iteration; the rich surrounding prey search strategy improves the surrounding prey method in the classic whale optimization algorithm, adding a randomly selected pool matrix member to balance the local search capability and the global search capability;

[0048] S5.5. Continuously update the position of the whale until the maximum number of iterations is reached and the global optimal solution is obtained; the global optimal solution is assigned to the corresponding hyperparameters of the gated recurrent unit neural network;

[0049] S5.6, inputting HF1 and HF2' of 80% cycle times of the battery selected in S5.2 into the neural network model with global optimal hyperparameters to obtain the predicted value of the battery capacity, and saving the optimal gated recurrent unit neural network capacity prediction model, and performing performance evaluation;

[0050] Performance evaluation methods include: R-squared (R 2 ), maximum absolute error (MAE), mean absolute error (MSE), and root mean square error (RMSE);

[0051]

[0052] In the formula, Cn' is the sampling value of the n'th cycle capacity; is the average of all cycle capacities of a battery; P n' is the estimated value of the n'th cycle capacity; m is the total number of cycles of a battery;

[0053] S6. Verify the generalizability of the prediction method.

[0054] Beneficial effects of the present invention:

[0055] The present invention proposes a lithium-ion battery capacity prediction method based on random charging fragments and indirect features. The charging time corresponding to two specific voltage ranges in the constant current charging process is selected as two characteristic parameters, and HF1 that can be collected through short-term charging data is used as an indirect feature. The prediction of HF2 is realized through Cox-Box transformation and extreme random forest algorithm. Then, HF2' is predicted by HF1 collected through short-term charging data, and a battery capacity prediction model is constructed based on enhanced whale optimization algorithm and gated recurrent unit neural network. Only the first 20% data of a certain battery is used for training, and the relative error of the capacity prediction result of this type of battery can be controlled within 1%, which greatly reduces the number of training sets while ensuring the prediction accuracy. On this basis, the capacity prediction values ​​of different types of batteries are obtained by using this method, which verifies that the constructed model has good prediction accuracy and generalization ability. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 A flow chart showing a method for predicting lithium-ion battery capacity based on random charging segments and indirect features;

[0057] Figure 2 The figure shows the relationship between the selected features HF1 and HF2 and the capacity aging at different cycle times; wherein, part (a) is the charging voltage curve of battery 1 at different cycle times; part (b) is the change graph of the health feature 1 of batteries 1-5; part (c) is the change graph of the capacity of batteries 1-5; part (d) is the change graph of the health feature 2 of batteries 1-5;

[0058] Figure 3 The result diagram of HF2 after training with 30% data randomly selected from battery 1; wherein, part (a) is the feature prediction diagram; part (b) is the relative error diagram;

[0059] Figure 4 It represents the prediction result diagram of the test set after training with HF1 and HF2' corresponding to the first 20% cycle number of battery 1; wherein, part (a) is the measurement value-prediction value result diagram; part (b) is the error result diagram;

[0060] Figure 5It shows the prediction result diagram of the test set after training with HF1 and HF2' corresponding to the first 20% cycle number of battery 2; wherein, part (a) is the measurement value-prediction value result diagram; part (b) is the error result diagram;

[0061] Figure 6 It shows the prediction result diagram of the test set after training with HF1 and HF2' corresponding to the first 20% cycle number of battery 3; wherein, part (a) is the measurement value-prediction value result diagram; part (b) is the error result diagram;

[0062] Figure 7 It shows the prediction result diagram of the test set after training with HF1 and HF2' using the first 20% cycle number of battery 4; wherein, part (a) is the measurement value-prediction value result diagram; part (b) is the error result diagram;

[0063] Figure 8 It shows the prediction result diagram of the test set after training with HF1 and HF2' using the first 20% cycle number of battery 5; wherein, part (a) is the measurement value-prediction value result diagram; part (b) is the error result diagram;

[0064] Fig. 9 A diagram showing the prediction results of the method applied to battery 6; wherein (a) part is the capacity prediction value; (b) part is the error diagram;

[0065] Fig.10 A diagram showing the prediction results of the method applied to battery 7; wherein (a) part is the capacity prediction value; and (b) part is the error diagram; DETAILED DESCRIPTION

[0066] In order to enable those skilled in the art to better understand the scheme of the present invention, the present invention is further described below in conjunction with the accompanying drawings and specific implementation methods.

[0067] A method for predicting lithium-ion battery capacity based on random charging segments and indirect features comprises the following steps:

[0068] S1. Perform cyclic charge and discharge tests on pre-set lithium-ion batteries of different models until the discharge capacity of each model of battery is less than 80% of the nominal capacity, and record the charge and discharge data in real time;

[0069] The pre-set different types of lithium-ion batteries are: different batteries of the same type and different batteries of different types;

[0070] In this embodiment, five soft-pack batteries with a rated capacity of 740 mAh and two B18650 cylindrical batteries with a rated capacity of 2550 mAh are selected as the test objects. The seven lithium batteries are named cell 1-7 respectively. The specific method flow chart is as follows: Figure 1 As shown;

[0071] The charging and discharging data include: charging voltage, discharging capacity and charging time;

[0072] S2. Obtain two health characteristics of the lithium-ion battery through two preset voltage intervals, and obtain a data set of battery health characteristics and battery capacity according to Spearman correlation analysis;

[0073] The two preset voltage intervals are obtained through the charging voltage curve of the soft pack battery cell 1 with a rated capacity of 740 mAh, as shown in FIG. Figure 2 As shown, the charging time when the voltage increases from 3.55V to 3.75V and from 3.80V to 4.15V is significantly reduced; the two preset voltage intervals are: 3.55V~3.75V and 3.80V~4.15V;

[0074] The two health features of lithium-ion batteries are health feature 1 (HF1) and health feature 2 (HF2). Health feature 1 (HF1) is regarded as an indirect feature to predict health feature 2 (HF2). Spearman correlation analysis is used to verify the correlation and obtain the health feature and capacity data set, as shown in Table 1.

[0075] Table 1: Spearman correlation coefficients for cells 1-5

[0076] Health characteristics Voltage range / V cell1 cell2 cell3 cell4 cell5 HF1 [3.55,3.75] 0.9988 0.9992 0.9994 0.9982 0.9982 HF2 [3.80,4.15] 0.9999 0.9999 0.9995 0.9998 0.9998

[0077] S3. Predict the health feature 2 (HF2) based on the extreme random forest algorithm to obtain the predicted feature (HF2'), the steps are as follows:

[0078] S3.1, HF1 and HF2 are subjected to Cox-Box transformation to enhance the linear correlation between HF1 and HF2, and HF10 and HF20 are obtained after the transformation; the calculation formula of Cox-Box transformation is as follows:

[0079]

[0080] In the formula, y i,k For each data point, we need y i,k >0,k=1,2,corresponding to HF1 and HF2;In(y i,k ) is for y i,k Take the logarithm; λ is a constant, and the maximum likelihood estimation method is needed to determine the λ value suitable for the data;

[0081] S3.2, randomly select HF10 and HF20 corresponding to 30% of the Cox-Box transformations, use HF10 as input and HF20 as output, and input them into the extreme random forest model for training; then input HF10 corresponding to the remaining 70% of the cycles into the trained extreme random forest model to obtain the predicted value HF20';

[0082] S3.3, perform Cox-Box inverse transformation on HF20' to obtain the prediction feature (HF2');

[0083] The calculation formula for the inverse Cox-Box transformation is as follows:

[0084]

[0085] In the formula, For Get the index;

[0086] S4, relative error analysis and Spearman correlation coefficient analysis were performed between the predicted feature (HF2') and the healthy feature 2 (HF2) to verify the prediction accuracy;

[0087] The expression of Spearman correlation coefficient analysis is as follows:

[0088]

[0089] In the formula, i=1, 2 represent HF1 and HF2 respectively; n represents the total number of samples, that is, the total number of HF or SOH; k represents the number of values; x represents HF; y represents the capacity; and Corresponding to the means of x and y respectively;

[0090] like Figure 3 As shown, only 3 predictions have relative errors exceeding 1%, and the maximum relative error is controlled within 2%, indicating good prediction accuracy; the Spearman correlation coefficient between HF2' and capacity is calculated to be 0.9989, and the Spearman correlation coefficient between HF2 and capacity is 0.9993, and the correlation is almost unchanged, which once again proves that the method has good prediction accuracy, and the predicted HF2' can be used as the feature of the final capacity prediction model;

[0091] S5, predicting the capacity of a single battery;

[0092] Specifically, the indirect features and predicted features are used as input, and the battery capacity is used as output, and are input into a gated recurrent unit neural network model based on an enhanced whale optimization algorithm for training. The training set is determined to be the data of the first 20% of the cycles, and the validation set is determined to be the data of the remaining 80% of the cycles, and the optimal capacity prediction model is saved;

[0093] The prediction steps are as follows:

[0094] S5.1. Set the network structure of the gated recurrent unit neural network and initialize the neural network hyperparameters;

[0095] S5.2, select any battery data in HF1 and the battery data corresponding to HF2' as the input of the gated recurrent unit neural network, use the capacity of the battery collected by S1 as the output of the neural network, determine that the training data is HF1, HF2' and capacity corresponding to the first 20% of the cycles, and the verification data is the data of the remaining 80% of the cycles;

[0096] S5.3, taking the mean square error of the capacity prediction value as the objective function, and defining the optimization objective as minimizing the mean square error of the capacity prediction value;

[0097] Specifically, the optimization object is set as the hyperparameters of the gated recurrent unit neural network, including: the number of iterations, the learning rate, and the number of neurons in the hidden layer; the enhanced whale optimization algorithm parameters are set, including: the number of whales, the maximum number of iterations, and the upper and lower limits of the optimization object;

[0098] In this embodiment, the optimization objects include: the number of neurons in the hidden layer, the upper and lower bounds of which are set to: [50, 500]; the initial learning rate, the upper and lower bounds of which are set to: [0.00001, 0.01]; the maximum number of iterations, the upper and lower bounds of which are set to: [150, 500]; L2 regularization, the upper and lower bounds of which are set to: [0.00001, 0.01]; Dropout ratio, the upper and lower bounds of which are set to: [0.1, 0.5];

[0099] S5.4. Evaluate the performance of the model based on the enhanced whale optimization algorithm;

[0100] Specifically, a group of whales (i.e., candidate solutions of the algorithm) are randomly initialized in the hyperparameter space; a gated recurrent unit neural network model is trained with the hyperparameter set represented by each whale, and the performance of the model is evaluated using the validation set;

[0101] The position of the whale (i.e., the optimal hyperparameter) is updated according to the rules of the enhanced whale optimization algorithm. The algorithm will simulate the predation behavior of whales, mainly including three stages: surrounding prey, spiral attack, and searching for prey, and optimize individuals to approach the optimal solution. Compared with the classic whale optimization algorithm, the enhanced whale optimization algorithm improves the performance of the classic whale optimization algorithm by introducing a pooling mechanism and three effective search strategies (migration search strategy, priority search strategy, and surrounding prey search strategy).

[0102] The pooling mechanism is defined as: given a matrix of size k, Pool = (p1, p2, ..., p k ), at the end of each iteration, the element P in the matrix Pool is generatedm =(p m,1 ,p m,2 ,...,p m,D ), wherein, in this embodiment, k=5;

[0103] Element P m The resulting formula is as follows:

[0104]

[0105] In the formula, Represents the element of the tth iteration; is the binary random vector in the tth iteration, is its inverse vector; is the worst solution in the tth iteration; It represents the optimal solution generated in the tth iteration A random position adjacent to the corresponding whale; where t=20 in this embodiment; at the end of each iteration, the pooling mechanism crosses the worst solution in the current iteration with a randomly generated promising solution to increase population diversity; this mechanism improves the algorithm's search capability by mixing the worst solution with the promising solution, avoiding premature convergence and local optimal traps;

[0106] The expression of the migration search strategy is as follows:

[0107]

[0108] Where: represents the position of the hth whale in the t+1th iteration; represents the random position of the whale in the search space in the tth iteration; rand is a random number between [0,1) and it follows a uniform distribution; δ max and δ min Corresponding to the upper and lower bounds of the optimization object respectively; δ best-max and δ best-min Then they correspond to the optimal solutions The migration search strategy increases exploration ability and population diversity by randomly separating a part of whales to cover unvisited areas. This strategy improves the algorithm's ability to escape local optimality by introducing randomness.

[0109] The expressions for the preferred search strategy are as follows:

[0110]

[0111] Where: represents the position of the hth whale in the tth iteration; and are positions 1 and 2 randomly selected from the matrix Pool in the tth iteration; MaxIt represents the maximum number of iterations set in the enhanced whale optimization algorithm; in order to improve the exploration ability of the classic whale optimization algorithm in searching for prey, the enhanced whale optimization algorithm adopts a priority selection search strategy, which discovers diverse solutions by dispersing whales in different areas of the search space, so a larger step size is required;

[0112] The expression of the encirclement prey search strategy is as follows:

[0113]

[0114] Where: is the position 3 randomly selected from the matrix Pool in the tth iteration; the rich surrounding prey search strategy improves the surrounding prey method in the classic whale optimization algorithm, adding a randomly selected pool matrix member to balance the local search capability and the global search capability;

[0115] S5.5. Continuously update the position of the whale until the maximum number of iterations is reached and the global optimal solution is obtained; the global optimal solution is assigned to the corresponding hyperparameters of the gated recurrent unit neural network;

[0116] S5.6, inputting HF1 and HF2' of 80% cycle times of the battery selected in S5.2 into the neural network model with global optimal hyperparameters to obtain the predicted value of the battery capacity, and saving the optimal gated recurrent unit neural network capacity prediction model, and performing performance evaluation;

[0117] Performance evaluation methods include: R-squared (R 2 ), maximum absolute error (MAE), mean absolute error (MSE), and root mean square error (RMSE);

[0118]

[0119] In the formula, C n' is the sampling value of the n'th cycle capacity; is the average of all cycle capacities of a battery; P n' is the estimated value of the n'th cycle capacity; m is the total number of cycles of a battery;

[0120] The results are as follows Figure 4 As shown, the solid line in the figure represents the actual capacity value, and the dotted line represents the estimated capacity value. From the results, it can be seen that the maximum error of the present invention is less than 0.98%;

[0121] S6. Verify the generalizability of the prediction method;

[0122] Specifically, for different batteries of the same type, the health characteristics of cell 2-5 obtained from the cyclic charge and discharge test in step 1 are directly input into the model saved in step 6 to obtain the capacity estimation value of the corresponding battery; the result is as follows: Figure 5 Parts (a) and (b) Figure 6 Parts (a) and (b) Figure 7 Parts (a) and (b) Figure 8 As shown in parts (a) and (b) and Table 2, the solid line in the figure represents the actual capacity value, and the dotted line represents the estimated capacity value. From the results, it can be seen that the R-squared (R 2 ), maximum absolute error (MAE), mean absolute error (MSE) and root mean square error (RMSE) are within 0.996, 0.265%, 0.443% and 0.344% respectively, which verifies that the proposed method only needs partial charging data of the power battery and only needs to train 20% of the data of a certain battery to accurately and stably predict the capacity of this type of battery, indicating the high efficiency and powerful estimation performance of the method, which makes it show broad application prospects in the field of battery health status prediction.

[0123] For different types of different batteries, extract all the data of cell 6 obtained from the cyclic charge and discharge test in step 1 as the model training set, repeat steps 2-5, and input cell 7 into the optimized model to obtain the estimated capacity value of the battery. Correspondingly, extract all the data of cell 7 obtained from the cyclic charge and discharge test in step 1 as the model training set, repeat steps 2-5, and input cell 6 into the optimized model to obtain the estimated capacity value of the battery. The results are as follows Fig. 9 Parts (a) and (b) Fig.10 As shown in parts (a) and (b) and Table 3, the solid line in the figure represents the actual capacity value, and the dotted line represents the estimated capacity value. It can be seen from the results that before the battery reaches the end of life (EOL), the maximum error of the present invention is less than 3.5%, and most of the prediction errors are within 2%, showing that the proposed method has good prediction accuracy and generalization.

[0124] Table 2: Prediction error of cells 2-5

[0125] cell <![CDATA[R 2 ]]> MAE(%) MSE(%) RMSE(%) 2 0.996 0.265 0.443 0.344 3 0.997 0.214 0.347 0.278 4 0.997 0.185 0.297 0.244 5 0.996 0.219 0.361 0.283

[0126] Table 3 Prediction errors of cell 6 and cell 7

[0127] cell <![CDATA[R 2 ]]> MAE(%) MSE(%) RMSE(%) 6 0.998 0.118 0.582 0.149 7 0.997 0.115 0.630 0.163

Claims

1. A lithium-ion battery capacity prediction method based on random charging segments and indirect features, characterized in that: The following steps are involved: S1. Perform cyclic charge and discharge tests on pre-set lithium-ion batteries of different models until the discharge capacity of each model of battery is less than 80% of the nominal capacity, and record the charge and discharge data in real time; The pre-set different types of lithium-ion batteries are: different batteries of the same type and different batteries of different types; The charging and discharging data include: charging voltage, discharging capacity and charging time; S2. Obtain two health characteristics of the lithium-ion battery through two preset voltage intervals, and obtain a data set of battery health characteristics and battery capacity according to Spearman correlation analysis; S3, predicting health feature 2 based on the extreme random forest algorithm to obtain the predicted feature; S4, relative error analysis and Spearman correlation coefficient analysis were performed on the predicted features and health features 2 to verify the prediction accuracy; S5, predicting the capacity of a single battery; S6. Verify the generalizability of the prediction method.

2. A lithium-ion battery capacity prediction method based on random charging segments and indirect features according to claim 1, characterized in that: The two preset voltage ranges are: 3.55V~3.75V and 3.80V~4.15V; The two health characteristics of lithium-ion batteries are: health characteristic 1 and health characteristic 2. By treating health characteristic 1 as an indirect characteristic, the prediction of health characteristic 2 can be achieved. Spearman correlation analysis was used to verify the correlation and obtain the health characteristics and capacity data sets.

3. A lithium-ion battery capacity prediction method based on random charging segments and indirect features according to claim 1, characterized in that: The steps of predicting the health feature 2 based on the extreme random forest algorithm to obtain the predicted feature are as follows: S3.

1. Perform Cox-Box transformation on HF1 and HF2 to obtain HF10 and HF20. The calculation formula of Cox-Box transformation is as follows: In the formula, y i,k For each data point, we need y i,k >0,k=1,2,corresponding to HF1 and HF2;In(y i,k ) is for y i,k Take the logarithm; λ is a constant, and the maximum likelihood estimation method is needed to determine the λ value suitable for the data; S3.2, randomly select HF10 and HF20 corresponding to 30% of the Cox-Box transformations, use HF10 as input and HF20 as output, and input them into the extreme random forest model for training; then input HF10 corresponding to the remaining 70% of the cycles into the trained extreme random forest model to obtain the predicted value HF20'; S3.3, perform Cox-Box inverse transformation on HF20' to obtain the prediction feature HF2'; The calculation formula for the inverse Cox-Box transformation is as follows: In the formula, For Take the index.

4. A lithium-ion battery capacity prediction method based on random charging segments and indirect features according to claim 1, characterized in that: The expression of the Spearman correlation coefficient analysis is as follows: In the formula, i=1, 2 represent HF1 and HF2 respectively; n represents the total number of samples; k represents the number of values; x represents HF; y represents the capacity; and Corresponding to the means of x and y respectively.

5. The method for predicting lithium-ion battery capacity based on random charging segments and indirect features according to claim 1, characterized in that: The steps of predicting the capacity of a single battery are as follows: S5.

1. Set the network structure of the gated recurrent unit neural network and initialize the neural network hyperparameters; S5.2, select any battery data in HF1 and the battery data corresponding to HF2' as the input of the gated recurrent unit neural network, use the capacity of the battery collected by S1 as the output of the neural network, determine that the training data is HF1, HF2' and capacity corresponding to the first 20% of the cycles, and the verification data is the data of the remaining 80% of the cycles; S5.3, taking the mean square error of the capacity prediction value as the objective function, and defining the optimization objective as minimizing the mean square error of the capacity prediction value; S5.

4. Evaluate the performance of the model based on the enhanced whale optimization algorithm; S5.

5. Continuously update the position of the whale until the maximum number of iterations is reached and the global optimal solution is obtained; the global optimal solution is assigned to the corresponding hyperparameters of the gated recurrent unit neural network; S5.6, inputting HF1 and HF2' of 80% cycle times of the battery selected in S5.2 into the neural network model with global optimal hyperparameters to obtain the predicted value of the battery capacity, and saving the optimal gated recurrent unit neural network capacity prediction model, and performing performance evaluation; Performance evaluation methods include: R-squared, maximum absolute error, mean absolute error, and root mean square error.

Images