Lithium battery SOH prediction method based on quantum particle swarm optimization hybrid neural network

Through quantum particle swarm optimization of hybrid neural networks, combined with multi-dimensional feature extraction and extrusion excitation attention-enhanced convolutional neural network, the problem of insufficient single-class features and poor generalization performance in lithium battery SOH prediction is solved, and the problem of further development is achieved. Accurate and stable health status prediction.

CN118980955AInactive Publication Date: 2024-11-19GUANGDONG UNIV OF TECH

Patent Information

Application Number
CN202411061250.6
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-05
Publication Date
2024-11-19
Estimated Expiration
Not applicable · inactive patent

AI Technical Summary

Technical Problem

The prior art has problems of insufficient single-class features and poor generalization performance in lithium battery SOH prediction, especially CNN-LSTM is difficult to effectively capture the dependencies and attention allocation between features, resulting in poor accuracy and generalization performance. .

Method used

Using a hybrid neural network method based on quantum particle swarm optimization, the characteristics are screened through multi-dimensional feature extraction and Pearson correlation coefficient analysis, combined with a convolutional neural network with enhanced extrusion excitation attention and a long and short-term memory network to predict lithium batteries SOH. And hyperparameter optimization is performed through quantum particle swarm to achieve global optimization.

Benefits of technology

A more accurate and effective prediction of the health status of lithium batteries is achieved, which improves the accuracy and generalization performance of the model, and reduces the generalization error and overfitting risks of the model.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118980955A_ABST
    Figure CN118980955A_ABST
Patent Text Reader

Abstract

The invention discloses a lithium battery SOH prediction method based on a quantum particle swarm optimization hybrid neural network. The lithium battery SOH prediction method comprises the following steps: S1, extracting multi-dimensional health features from two groups of same batteries, and performing Pearson's correlation coefficient analysis to obtain features with high correlation coefficients; s2, taking the data of one group of batteries as training set data, and taking the data of the other group of batteries as test set data; s3, inputting the training set data into a hybrid neural network for prediction, namely, firstly inputting the convolutional neural network for extruding and exciting attention enhancement to fully extract features, and then fully extracting a long-term dependency relationship by using a long-short-term memory neural network; and finally, performing hyper-parameter optimization by using a quantum particle swarm so as to achieve global optimization. And S4, inputting the test set into the hybrid neural network based on quantum particle swarm optimization, predicting the health state of the lithium battery, and performing error analysis.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to battery health prediction technology, in particular to a lithium battery SOH prediction method based on quantum particle swarm optimization hybrid neural network. Background Art

[0002] The battery management system (BMS) is an important component of new energy vehicles, and the battery's state of health (SOH) is one of the important state quantities that the battery management system pays attention to. Lithium-ion batteries have become energy storage devices for electric vehicles due to their high energy density, long cycle life, and low pollution. However, lithium-ion batteries will age after multiple charge and discharge cycles, leading to problems such as leakage and local short circuits, and eventually causing fires or even explosions. Therefore, it is very important to accurately estimate the SOH of lithium batteries. However, due to the complex and coupled internal aging mechanism of the battery, accurately estimating its SOH is a challenge.

[0003] In order to quantitatively describe the battery decay process, this article defines the SOH definition of lithium-ion batteries using battery capacity as follows:

[0004]

[0005] Among them, "Q C ” and “ Q. R " are the current capacity and rated capacity of the battery respectively. When SOH < 0.8, the lithium battery reaches the aging and retirement standard.

[0006] There are three methods for estimating SOH: experimental method, model-based method, and data-driven method. The experimental method has low computational complexity, but online operation requires additional hardware support and has poor noise resistance. Model-based methods include electrochemical models and equivalent circuit models. Due to the complexity of the electrochemical mechanism of lithium batteries, it is difficult to build a battery aging model. The use of an equivalent circuit model can simplify the original battery circuit and make it easier to measure. However, the accuracy of the estimation is not satisfactory. The data-driven method estimates SOH because it does not involve the complex internal mechanism of the battery. It only needs to extract health features (HFs) from the battery status data, such as voltage, current, temperature, etc., and use machine learning or deep learning algorithms to fit the nonlinear mapping relationship between HFs and SOH to achieve SOH estimation. It has strong flexibility and scalability.

[0007] For the prediction of battery SOH using data-driven methods, it is necessary to extract typical features from the battery status data and establish a mapping relationship between them and the health status. The existing technology has the following problems: 1. Currently, single-class features are generally used for battery SOH prediction, which cannot fully characterize the battery aging process, resulting in limited accuracy of the model; 2. Ordinary CNN-LSTM cannot effectively capture the dependency between features and allocate attention according to the importance of features; when the parameters of the data-driven model cannot be flexibly set according to the characteristic data of the battery, the generalization performance of the model is poor. Summary of the invention

[0008] In view of the above-mentioned defects, the purpose of the present invention is to propose a lithium battery SOH prediction method based on quantum particle swarm optimization hybrid neural network, which can extract and analyze multi-dimensional features of the battery and predict the health status of the lithium battery more accurately and effectively.

[0009] To achieve this object, the present invention adopts the following technical solutions:

[0010] A lithium battery SOH prediction method based on quantum particle swarm optimization hybrid neural network includes the following steps:

[0011] S1. Extract multi-dimensional health features from two sets of the same battery and obtain features with high correlation coefficients through Pearson correlation coefficient analysis;

[0012] S2, using the data of one group of batteries as training set data, and the data of another group of batteries as test set data;

[0013] S3, input the training set data into the hybrid neural network for prediction, that is, first input the squeeze-excited attention-enhanced convolutional neural network to fully extract features, and then use the long short-term memory neural network to fully extract long-term dependencies; finally, use the quantum particle swarm to optimize the hyperparameters to achieve the global optimum;

[0014] S4. Input the test set into the hybrid neural network based on quantum particle swarm optimization to predict the SOH of the lithium battery and perform error analysis.

[0015] Further, in step S1:

[0016] The method comprises step S11, performing multi-dimensional feature extraction on the battery, including extracting 12 health features closely related to the SOH of the lithium battery in seven dimensions, namely, voltage, current, temperature, incremental capacity, differential thermovoltammetry, energy and sample entropy, which are the initial time point of constant voltage charging, the area covering the voltage curve during constant current charging, the constant current charging time, the area covered by the current curve during constant current charging, the ratio of the maximum temperature difference to time in constant current discharge, the area covered by the temperature curve during discharge, the time corresponding to the peak value of the incremental capacity curve, the voltage corresponding to the peak value of the incremental capacity curve, the second trough position of the differential thermovoltammetry curve, the second trough of the differential thermovoltammetry curve, the energy in the constant current discharge stage and the sample entropy of the discharge voltage;

[0017] Step S12, performing Pearson correlation coefficient analysis on the above 12 health characteristics and the battery SOH value:

[0018]

[0019] In formula (2), ρ is the correlation coefficient, F is the health characteristic, σ F With σ SOH are the standard deviations of feature F and SOH, respectively. Healthy features with a Pearson correlation coefficient higher than 85% are selected as training samples and test samples.

[0020] Furthermore, in step S3, the convolutional neural network is hybridly connected with the squeeze-stimulation attention mechanism to obtain a convolutional neural network enhanced based on squeeze-stimulation attention, and is connected to a long short-term memory network to obtain a hybrid neural network.

[0021] Furthermore, step S3 is specifically as follows:

[0022] The main execution process of the quantum particle swarm algorithm is as follows:

[0023] S31. Population initialization

[0024] Set the number of iterations t = 0, and initialize each particle of the quantum particle swarm in the M-dimensional space to be solved, and randomly obtain the spatial position X of each particle i (0), and initialize the optimal position of each particle to P i (0) = X i (0);

[0025] S32, calculate the average optimal position of particles mbest according to formula (3), where M is the population size, P i is the optimal position of individual particles;

[0026]

[0027] S33, calculate the particle fitness value, and the particle's own optimal position pbest Update; the update method is to change X i The fitness value of (t) and P i (t-1) fitness value comparison, if X i The fitness value of (t) is better than P i (t-1) fitness value, let P i (t) = X i (t), otherwise P i (t) = P i (t-1);

[0028] S34. For particle i, compare its P i (t) and the current global optimal position P g (t-1) fitness value, if P i The fitness value of (t) is better than P g (t-1) fitness value, let P g (t) = P i (t), otherwise P g (t) = P g (t-1);

[0029] S35, obtaining a random position of each particle of the particle group;

[0030] S36, updating the latest position of the particle according to formula (4);

[0031]

[0032] Where P id (t) and P gd (t) are the individual optimal position and global optimal value of particle d, φ and μ are random numbers between [0,1], and β is the contraction and expansion coefficient; id (t+1) and x id (t) represents the position of particle i in the dth dimension at iterations t+1 and t, respectively.

[0033] Preferably, the step S4 is specifically as follows:

[0034] The evaluation indicators commonly used in lithium battery SOH prediction to evaluate regression prediction results and measure the accuracy of prediction models are selected, namely, the fitting coefficient R2 (Coefficient of Determination, also called R Squared), the mean absolute percentage error MAPE (Mean Absolute Percentage Error), the root mean square error RMSE (Root Mean Squared Error) and the mean absolute error MAE (Mean Absolute Error), and the formulas are:

[0035]

[0036] In the formula, n is the number of samples, y is i is the actual sample value, is the sample mean, is the predicted sample value.

[0037] Preferably, in step S11:

[0038] Voltage related characteristics:

[0039] The starting point of constant voltage charging advances as the number of cycles increases; the coverage area of ​​the voltage curve during constant current charging decreases as the number of cycles increases. Therefore, the starting point of constant voltage charging (F1) and the coverage area of ​​the voltage curve during constant current charging (F2) can be extracted as voltage-related features;

[0040] Current related characteristics:

[0041] The constant current charging time decreases as the number of cycles increases, and the coverage area of ​​the current curve during the constant current charging process decreases as the number of cycles increases. Therefore, the constant current charging time (F3) and the coverage area of ​​the current curve during the constant current charging process (F4) can be used as current-related features in advance;

[0042] Temperature-related characteristics:

[0043] During constant current discharge, the time when the battery reaches the maximum temperature advances with the increase of the number of cycles. This paper extracts the ratio of the maximum temperature difference to time during constant current discharge (F5) and the area covered by the temperature curve during discharge (F6) as temperature-related features;

[0044] Incremental capacity related features:

[0045] Incremental capacity analysis is a powerful non-invasive electrochemical analysis method that can explore subtle changes in electrochemical processes caused by capacity loss; the IC curve is the change curve of IC with discharge voltage, and the calculation process is as follows:

[0046]

[0047] Since the collected data has certain noise, it is necessary to pre-process the data. The IC curve is filtered by the SG filtering method. The position of the peak of the IC curve advances with the increase of the number of cycles, and the voltage value corresponding to the peak of the IC curve decreases with the increase of the number of cycles. Therefore, the position of the peak of the IC curve and the voltage value corresponding to the peak of the IC curve are extracted as IC-related features, the time corresponding to the peak of the incremental capacity curve (F7), and the voltage corresponding to the peak of the incremental capacity curve (F8);

[0048] Differential thermovoltammetry related features:

[0049] Differential thermovoltammetry is a signal analysis technique that can mine information related to battery degradation through changes in entropy; it determines the state of the battery by calculating the difference between temperature and voltage; DTV is calculated as follows:

[0050]

[0051] Where dT represents the differential value of temperature, and dV represents the differential value of voltage;

[0052] The DTV curve still has a lot of noise. Similarly, the SG filtering method can be used to filter the original data. The position of the second trough of the DTV curve (F9) and the second trough value of the DTV curve (F10) are extracted as DTV related features.

[0053] Energy-related characteristics:

[0054] For different numbers of cycles, the time required for the battery to complete the constant current charging process is different, and the time required to complete the constant current discharging process is also different, resulting in different amounts of energy absorbed or released in each cycle;

[0055] The energy-based method can calculate the constant current charging or discharging phase energy according to formula (12) to characterize the battery aging process.

[0056]

[0057] Where, respectively, are the start time and end time of charging or discharging, and respectively are the voltage and current at cycle i;

[0058] Therefore, the energy (F11) in the constant current discharge stage can be extracted as an energy-related feature;

[0059] Sample entropy related features:

[0060] Sample entropy is a statistic used to measure the complexity of time series and plays an important role in analyzing signal complexity and predictability. As the battery decays, the measurable voltage response under the same stimulus will also change accordingly. The salient feature of this method is that at the end of each discharge cycle, the SOH is estimated by simply calculating the sample entropy of the voltage, which does not involve a complex calculation process and avoids the cumulative error caused by integration. Based on the above advantages, the sample entropy feature is widely used to describe the battery aging process. Suppose a time series X = {x(1), x(2), ..., x(N)} with a length of N, the calculation method of sample entropy is as follows:

[0061] Step1: Construct the time series X into an m-dimensional vector, i.e.,

[0062] X(i) = {x(i), x(i + 1),.., x(i + m - 1)} (13)

[0063] where i = 1, 2,..., N - m + 1. X(i) represents a subsequence constructed from the original time series X or is called a vector; specifically, X(i) is a subsequence containing m consecutive data points, starting from the i-th data point of the original time series X;

[0064] Step2: Define the distance between vectors as when i is not equal to j, it is the largest of the differences between their corresponding elements, i.e.:

[0065]

[0066] where d[X(i), X(j)] is the largest distance among the differences between the corresponding elements of the two subsequences X(i) and X(j). x(i + k) is the (k + 1)-th element in the subsequence X(i), and x(j + k) is the (k + 1)-th element in the subsequence X(j), because the index usually starts from 0 or 1;

[0067] Step3: Given a threshold r (r > 0), count the number of < r and take the ratio with the total number of vectors N - m, i.e.,

[0068]

[0069] where is a ratio value, representing the ratio of the number of m-dimensional vectors similar (distance less than r) to the i-th m-dimensional vector X(i) in the time series to the total number of m-dimensional vectors. Here, r is a preset threshold used to define the "similarity" between two vectors. N is the length of the original time series X, i.e., the number of data points in the time series; m is the embedding dimension or called the pattern length, which determines the length of each subsequence (or called vector) constructed from the original time series.

[0070] Step4: Take the average of all the results obtained from Equation (8), i.e.:

[0071]

[0072] where B m (r) is an average ratio value, representing the average ratio of the vectors among all m-dimensional vectors that are similar (i.e., the distance between them is less than the preset threshold r) to at least one other m-dimensional vector.

[0073] Step5: Then increase the dimension m by 1 and repeat Step1 - 4;

[0074] Step 6: Theoretically, the sample entropy of this sequence is:

[0075]

[0076] But in reality, N cannot be infinite, but a finite value, so the estimated value of sample entropy is:

[0077]

[0078] The sample entropy (F12) is extracted as a feature.

[0079] One of the above technical solutions includes the following beneficial effects: 1. In view of the limitations of health feature extraction, the present invention extracts health features from multiple dimensions to more comprehensively characterize the battery aging process to improve the accuracy of the model. 2. In view of the randomness of attention allocation in deep models, the present invention proposes a hybrid neural network based on squeeze-stimulated attention enhancement to accurately predict battery SOH. 3. The present invention adopts a parameter optimization method to find the most suitable hyperparameters for deep learning models. BRIEF DESCRIPTION OF THE DRAWINGS

[0080] Figure 1 It is a schematic diagram of a lithium battery SOH prediction model based on a quantum particle swarm optimization hybrid neural network according to an embodiment of the present invention;

[0081] Figure 2 It is a schematic diagram of specific steps of data preprocessing according to an embodiment of the present invention;

[0082] Figure 3 is a schematic diagram of a hybrid neural network structure of an embodiment of the present invention;

[0083] Figure 4 is a prediction flow chart based on quantum particle swarm optimization hybrid neural network according to an embodiment of the present invention;

[0084] Figure 5 is a schematic diagram of a squeeze-stimulated attention structure of an embodiment of the present invention;

[0085] Figure 6 is a battery capacity decay curve of an embodiment of the present invention;

[0086] Figure 7 1 is a voltage curve of different cycles of an embodiment of the present invention;

[0087] Figure 8 1 is a current curve of different cycles of an embodiment of the present invention;

[0088] Fig. 9is a temperature curve of different cycles of an embodiment of the present invention;

[0089] Fig.10 1 is an IC curve of different cycles of an embodiment of the present invention;

[0090] Fig.11 are DTV curves of different cycles of one embodiment of the present invention;

[0091] Fig.12 is a curve showing the relationship between energy characteristics and SOH of an embodiment of the present invention;

[0092] Fig.13 is a curve showing the relationship between the sample entropy feature and SOH according to an embodiment of the present invention;

[0093] Fig.14 This is the SOH curve fitting effect of an embodiment of the present invention. DETAILED DESCRIPTION

[0094] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and cannot be understood as limiting the present invention.

[0095] like Figure 1 As shown, a lithium battery SOH prediction method based on quantum particle swarm optimization hybrid neural network is characterized by comprising the following steps:

[0096] S1. Extract multi-dimensional health features from two sets of the same battery and obtain features with high correlation coefficients through Pearson correlation coefficient analysis;

[0097] S2, using the data of one group of batteries as training set data, and the data of another group of batteries as test set data;

[0098] S3, input the training set data into the hybrid neural network for prediction, that is, first input the squeeze-excited attention-enhanced convolutional neural network to fully extract features, and then use the long short-term memory neural network to fully extract long-term dependencies; finally, use the quantum particle swarm to optimize the hyperparameters to achieve the global optimum;

[0099] S4. Input the test set into the quantum particle swarm optimization hybrid neural network (QPSO-CL-SENet) to predict the SOH of the lithium battery and perform error analysis.

[0100] This method combines a variety of advanced technologies including multi-dimensional feature extraction and analysis, construction of hybrid neural networks, quantum particle swarms for hyperparameter optimization, and error analysis. It can more accurately and effectively predict the state of health (SOH) of lithium batteries, which is of great significance for extending battery life and improving the performance of battery management systems.

[0101] like Figure 2 As shown, in step S1:

[0102] The method comprises step S11, performing multi-dimensional feature extraction on the battery, including extracting 12 health features closely related to the SOH of the lithium battery in seven dimensions, namely, voltage, current, temperature, incremental capacity, differential thermovoltammetry, energy and sample entropy, which are the initial time point of constant voltage charging, the area covering the voltage curve during constant current charging, the constant current charging time, the area covered by the current curve during constant current charging, the ratio of the maximum temperature difference to time in constant current discharge, the area covered by the temperature curve during discharge, the time corresponding to the peak value of the incremental capacity curve, the voltage corresponding to the peak value of the incremental capacity curve, the second trough position of the differential thermovoltammetry curve, the second trough of the differential thermovoltammetry curve, the energy in the constant current discharge stage and the sample entropy of the discharge voltage;

[0103] Step S12, performing Pearson correlation coefficient analysis on the above 12 health characteristics and the battery SOH value:

[0104]

[0105] In formula (2), ρ is the correlation coefficient, F is the health characteristic, σ F With σ SOH are the standard deviations of feature F and SOH, respectively. Healthy features with a Pearson correlation coefficient higher than 85% were selected as training samples and test samples. Through Pearson correlation coefficient analysis, features with high correlation were selected for modeling, which reduced redundant features and improved the efficiency and accuracy of the model.

[0106] like Figure 3 and Figure 5 As shown, in step S3, the convolutional neural network (CNN) is mixed and connected with the squeeze-excited attention mechanism (SENet) to obtain a convolutional neural network based on squeeze-excited attention enhancement, and connected to LSTM (long short-term memory network) to obtain a hybrid neural network CNN-LSTM-SENet. By combining the convolutional neural network (CNN) and the long short-term memory network (LSTM), the features and long-term dependencies in the battery data can be fully extracted, thereby improving the accuracy of the prediction.

[0107] like Figure 4 As shown, step S3 is specifically as follows:

[0108] The main execution process of the quantum particle swarm algorithm is as follows:

[0109] S31. Population initialization

[0110] Set the number of iterations t = 0, and initialize each particle of the quantum particle swarm in the M-dimensional space to be solved, and randomly obtain the spatial position X of each particle i (0), and initialize the optimal position of each particle to P i (0) = X i (0);

[0111] S32, calculate the average optimal position of particles mbest according to formula (3), where M is the population size, P i is the optimal position of individual particles;

[0112]

[0113] S33, calculate the particle fitness value, and the particle's own optimal position p best Update; the update method is to change X i The fitness value of (t) and P i (t-1) fitness value comparison, if X i The fitness value of (t) is better than P i (t-1) fitness value, let P i (t) = X i (t), otherwise P i (t) = P i (t-1);

[0114] S34. For particle i, compare its P i (t) and the current global optimal position P g (t-1) fitness value, if P i The fitness value of (t) is better than P g (t-1) fitness value, let P g (t) = P i (t), otherwise P g (t) = P g (t-1);

[0115] S35, obtaining a random position of each particle of the particle group;

[0116] S36, updating the latest position of the particle according to formula (4);

[0117]

[0118] Where P id (t) and P gd(t) are the individual optimal position and global optimal value of particle d, φ and μ are random numbers between [0,1], and β is the contraction and expansion coefficient; id (t+1) and x id (t) represents the position of particle i in the dth dimension at iterations t+1 and t, respectively. Quantum particle swarm optimization (QPSO) is used for hyperparameter optimization, which can effectively avoid local optimal problems, find the global optimal solution, and improve the performance of the model.

[0119] In addition, the step S4 is specifically as follows:

[0120] The evaluation indicators commonly used in lithium battery SOH prediction to evaluate regression prediction results and measure the accuracy of prediction models are selected, namely, the fitting coefficient R2 (Coefficient of Determination, also called R Squared), the mean absolute percentage error MAPE (Mean Absolute Percentage Error), the root mean square error RMSE (Root Mean Squared Error) and the mean absolute error MAE (Mean Absolute Error), and the formulas are:

[0121]

[0122]

[0123] In the formula, n is the number of samples, y is i is the actual sample value, is the sample mean, is the predicted sample value.

[0124] After the prediction results are obtained, error analysis can help identify the deficiencies of the model, further optimize the model, and improve the reliability of the prediction. The prediction accuracy of the model can be evaluated by the size of these evaluation indicators.

[0125] Using the content from step S11:

[0126] The present invention predicts the battery health value and proves the effectiveness and accuracy of the model.

[0127] Using NASA's lithium-ion battery aging data set, the capacity decay curves of batteries B0005, B0006, and B0007 are as follows: Figure 6 As shown in Table 1, different dimensions of health features are extracted from the charge and discharge curves;

[0128]

[0129]

[0130] Table 1

[0131] Voltage related characteristics:

[0132] like Figure 7 As shown in the figure, the starting point of constant voltage charging advances with the increase of cycle number; the coverage area of ​​the voltage curve of constant current charging process decreases with the increase of cycle number. Therefore, the starting point of constant voltage charging (F1) and the coverage area of ​​the voltage curve of constant current charging process (F2) can be extracted as voltage-related features;

[0133] Current related characteristics:

[0134] like Figure 8 As shown, the constant current charging time decreases with the increase of the number of cycles, and the coverage area of ​​the current curve of the constant current charging process decreases with the increase of the number of cycles. Then, the constant current charging time (F3) and the coverage area of ​​the current curve of the constant current charging process (F4) can be used as current-related features in advance;

[0135] Temperature-related characteristics:

[0136] like Fig. 9 As shown in the figure, during the constant current discharge process, the time when the battery reaches the maximum temperature is advanced as the number of cycles increases. This paper extracts the ratio of the maximum temperature difference to time in the constant current discharge (F5) and the area covered by the temperature curve during the discharge process (F6) as temperature-related features;

[0137] Incremental capacity related features:

[0138] Incremental capacity analysis is a powerful non-invasive electrochemical analysis method that can explore subtle changes in electrochemical processes caused by capacity loss; the IC curve is the change curve of IC with discharge voltage, and the calculation process is as follows:

[0139]

[0140] like Fig.10 As shown in the figure, since the collected data has certain noise, it is necessary to pre-process the data. The IC curve is filtered by the SG filtering method. The position of the peak of the IC curve is advanced as the number of cycles increases, and the voltage value corresponding to the peak of the IC curve decreases as the number of cycles increases. Therefore, the position of the peak of the IC curve and the voltage value corresponding to the peak of the IC curve are extracted as the time corresponding to the peak of the incremental capacity curve (F7) and the voltage corresponding to the peak of the incremental capacity curve (F8) related to the IC characteristics.

[0141] Differential thermovoltammetry related features:

[0142] Differential thermovoltammetry is a signal analysis technique that can mine information related to battery degradation through changes in entropy; it determines the state of the battery by calculating the difference between temperature and voltage; DTV is calculated as follows:

[0143]

[0144] Where dT represents the differential value of temperature, and dV represents the differential value of voltage;

[0145] The DTV curve still has a lot of noise. Similarly, the SG filtering method can be used to filter the original data. The DTV curve after filtering is as follows: Fig.11 As shown, the position of the second trough of the DTV curve (F9) and the second trough value of the DTV curve (F10) are extracted as DTV related features;

[0146] Energy-related characteristics:

[0147] For different numbers of cycles, the time required for the battery to complete the constant current charging process is different, and the time required to complete the constant current discharging process is also different, resulting in different amounts of energy absorbed or released in each cycle;

[0148] The energy-based method can calculate the constant current charging or discharging phase energy according to formula (12) to characterize the battery aging process.

[0149]

[0150] Where, respectively, are the start time and end time of charging or discharging, and respectively are the voltage and current at cycle i;

[0151] Therefore, the energy (F11) in the constant current discharge stage can be extracted as the energy-related feature. The relationship between the energy feature and SOH is as follows: Fig.12 As shown;

[0152] Sample entropy related features:

[0153] Sample entropy is a statistic used to measure the complexity of time series and plays an important role in analyzing signal complexity and predictability. As the battery decays, the measurable voltage response under the same stimulus will also change accordingly. The salient feature of this method is that at the end of each discharge cycle, the SOH is estimated by simply calculating the sample entropy of the voltage, which does not involve a complex calculation process and avoids the cumulative error caused by integration. Based on the above advantages, the sample entropy feature is widely used to describe the battery aging process. Suppose a time series X = {x(1), x(2), ..., x(N)} with a length of N, the calculation method of sample entropy is as follows:

[0154] Step1: Construct the time series X into an m-dimensional vector, i.e.,

[0155] X(i) = {x(i), x(i + 1),..., x(i + m - 1)} (13)

[0156] where i = 1, 2,..., N - m + 1. X(i) represents a subsequence constructed from the original time series X or is called a vector; specifically, X(i) is a subsequence containing m consecutive data points, starting from the i-th data point of the original time series X;

[0157] Step2: Define the distance between vectors as the largest difference among the corresponding elements when i is not equal to j, i.e.:

[0158]

[0159] where d[X(i), X(j)] is the largest distance among the differences between the corresponding elements of the two subsequences X(i) and X(j). x(i + k) is the (k + 1)-th element in the subsequence X(i), and x(j + k) is the (k + 1)-th element in the subsequence X(j), as the index usually starts from 0 or 1;

[0160] Step3: Given a threshold r (r > 0), count the number of < r and calculate its ratio to the total number of vectors N - m, i.e.,

[0161]

[0162] where is a ratio value, representing the ratio of the number of m-dimensional vectors that are similar (distance less than r) to the i-th m-dimensional vector X(i) in the time series to the total number of m-dimensional vectors. Here, r is a preset threshold used to define the "similarity" between two vectors. N is the length of the original time series X, i.e., the number of data points in the time series; m is the embedding dimension or the pattern length, which determines the length of each subsequence (or vector) constructed from the original time series.

[0163] Step4: Take the average of all the results obtained from Equation (8), i.e.:

[0164]

[0165] where B m (r) is an average ratio value, representing the average ratio of the vectors among all m-dimensional vectors that are similar (i.e., the distance between them is less than the preset threshold r) to at least one other m-dimensional vector.

[0166] Step5: Then increase the dimension m by 1 and repeat Steps 1 - 4;

[0167] Step 6: Theoretically, the sample entropy of this sequence is:

[0168]

[0169] But in reality, N cannot be infinite, but a finite value, so the estimated value of sample entropy is:

[0170]

[0171] The relationship between the sample entropy characteristics of the discharge voltage and SOH is as follows: Fig.13 As shown, the sample entropy (F12) is extracted as a feature.

[0172] Health features are extracted from multiple dimensions of battery status. These multi-dimensional features more comprehensively characterize the battery capacity decay process. Features of different dimensions can complement each other, reduce the errors that may be caused by a single feature, and thus improve the accuracy of prediction. Multi-dimensional features can increase data diversity, reduce the risk of model overfitting, and make the model more stable on different data sets.

[0173] In order to analyze the correlation between multidimensional features and SOH, this paper uses Pearson correlation analysis, and the Pearson correlation coefficient is shown in Table 2. Since the correlation coefficient between F10 and SOH is less than 0.85, it is eliminated.

[0174]

[0175] Table 2 Pearson correlation coefficients between multidimensional features and SOH

[0176] The prediction example of the lithium battery health status prediction model based on quantum particle swarm optimization hybrid neural network uses B0005, B0006 and B0007 batteries in the NASA data set to verify the accuracy of the developed method. One of the batteries (for example, B0007) is used as the training set, and the other battery (B0006) is used as the test set. The population size of the quantum particle swarm is set to 20 and the number of iterations is set to 10; Fig.14 The SOH curve fitting effect is shown with B7 as the training set and B6 as the test set. Table 3 shows the SOH prediction comparison results under different methods. Extensive experimental verification and comparative studies show that compared with the prediction effects of LSTM, CNN-LSTM, and CNN-LSTM-SENet, the SOH curve fitting effect of the proposed method is better and the evaluation index is better. Compared with the optimal baseline model CNN-LSTM-SENet, the prediction error RMSE of the proposed method is reduced by 0.29%, MAE is reduced by 0.17%, MAPE is reduced by 0.16%, and R^2 is increased by 0.61%.

[0177]

[0178] Table 3 Comparison of SOH prediction results under different methods

[0179] In the full text, Convolutional Neural Networks (CNN); Squeeze-and-Excitation Networks (SENet); Long Short-Term Memory Network (LSTM); Hybrid Neural Network (CNN-LSTM-SENet); Hybrid Neural Network based on quantum particle swarm optimization (QPSO-CL-SENet); State of Health (SOH).

[0180] The technical principle of the present invention is described above in conjunction with specific embodiments. These descriptions are only for explaining the principle of the present invention and cannot be interpreted as limiting the scope of protection of the present invention in any way. Based on the explanations herein, those skilled in the art can associate other specific implementations of the present invention without paying creative labor, and these methods will fall within the scope of protection of the present invention.

Claims

1. A lithium battery SOH prediction method based on quantum particle swarm optimization hybrid neural network, characterized in that: The steps include: S1. Extract multi-dimensional health features from two sets of the same battery and obtain features with high correlation coefficients through Pearson correlation coefficient analysis; S2, using the data of one group of batteries as training set data, and the data of another group of batteries as test set data; S3, input the training set data into the hybrid neural network for prediction, that is, first input the squeeze-excited attention-enhanced convolutional neural network to fully extract features, and then use the long short-term memory neural network to fully extract long-term dependencies; finally, use the quantum particle swarm to optimize the hyperparameters to achieve the global optimum; S4. Input the test set into the hybrid neural network based on quantum particle swarm optimization to predict the SOH of the lithium battery and perform error analysis.

2. The lithium battery SOH prediction method based on quantum particle swarm optimization hybrid neural network according to claim 1 is characterized in that: In step S1: The method comprises step S11, performing multi-dimensional feature extraction on the battery, including extracting 12 health features closely related to the SOH of the lithium battery in seven dimensions, namely, voltage, current, temperature, incremental capacity, differential thermovoltammetry, energy and sample entropy, which are the initial time point of constant voltage charging, the area covering the voltage curve during constant current charging, the constant current charging time, the area covered by the current curve during constant current charging, the ratio of the maximum temperature difference to time in constant current discharge, the area covered by the temperature curve during discharge, the time corresponding to the peak of the incremental capacity curve, the voltage corresponding to the peak of the incremental capacity curve, the second trough position of the differential thermovoltammetry curve, the second trough of the differential thermovoltammetry curve, the energy in the constant current discharge stage and the sample entropy of the discharge voltage; Step S12, performing Pearson correlation coefficient analysis on the above 12 health characteristics and the battery SOH value: In formula (2), ρ is the correlation coefficient, F is the health characteristic, σ F With σ SOH are the standard deviations of feature F and SOH, respectively. Healthy features with a Pearson correlation coefficient higher than 85% are selected as training samples and test samples.

3. The lithium battery SOH prediction method based on quantum particle swarm optimization hybrid neural network according to claim 2 is characterized in that: In step S3, the convolutional neural network is hybridly connected with the squeeze-stimulated attention mechanism to obtain a convolutional neural network enhanced based on squeeze-stimulated attention, and is connected to a long short-term memory network to obtain a hybrid neural network.

4. The lithium battery SOH prediction method based on quantum particle swarm optimization hybrid neural network according to claim 3 is characterized in that: Step S3 is specifically as follows: The main execution process of the quantum particle swarm algorithm is as follows: S31. Population initialization Set the number of iterations t = 0, and initialize each particle of the quantum particle swarm in the M-dimensional space to be solved, and randomly obtain the spatial position X of each particle i (0), and initialize the optimal position of each particle to P i (0) = X i (0); S32, calculate the average optimal position of particles mbest according to formula (3), where M is the population size, P i is the optimal position of individual particles; S33, calculate the particle fitness value, and the particle's own optimal position p best Update; the update method is to change X i The fitness value of (t) and P i (t-1) fitness value comparison, if X i The fitness value of (t) is better than P i (t-1) fitness value, let P i (t) = X i (t), otherwise P i (t) = P i (t-1); S34. For particle i, compare its P i (t) and the current global optimal position P g (t-1) fitness value, if P i The fitness value of (t) is better than P g (t-1) fitness value, let P g (t) = P i (t), otherwise P g (t) = P g (t-1); S35, obtaining a random position of each particle of the particle group; S36, updating the latest position of the particle according to formula (4); Where P id (t) and P gd (t) are the individual optimal position and global optimal value of particle d, φ and μ are random numbers between [0,1], and β is the contraction and expansion coefficient; id (t+1) and x id (t) represents the position of particle i in the dth dimension at iterations t+1 and t, respectively.

5. The lithium battery SOH prediction method based on quantum particle swarm optimization hybrid neural network according to claim 4 is characterized in that: The step S4 is specifically as follows: The evaluation indicators commonly used in lithium battery SOH prediction to evaluate regression prediction results and measure the accuracy of prediction models are selected, namely, the fitting coefficient R2 (Coefficient of Determination, also called R Squared), the mean absolute percentage error MAPE (Mean Absolute Percentage Error), the root mean square error RMSE (Root Mean Squared Error) and the mean absolute error MAE (Mean Absolute Error), and the formulas are: In the formula, n is the number of samples, y is i is the actual sample value, is the sample mean, is the predicted sample value.

6. The lithium battery SOH prediction method based on quantum particle swarm optimization hybrid neural network according to claim 5 is characterized in that: In step S11: Voltage related characteristics: The starting point of constant voltage charging advances as the number of cycles increases; the coverage area of ​​the voltage curve during constant current charging decreases as the number of cycles increases. Therefore, the starting point of constant voltage charging (F1) and the coverage area of ​​the voltage curve during constant current charging (F2) can be extracted as voltage-related features; Current related characteristics: The constant current charging time decreases as the number of cycles increases, and the coverage area of ​​the current curve during the constant current charging process decreases as the number of cycles increases. Therefore, the constant current charging time (F3) and the coverage area of ​​the current curve during the constant current charging process (F4) can be used as current-related features in advance; Temperature related characteristics: During constant current discharge, the time when the battery reaches the maximum temperature advances with the increase of the number of cycles. This paper extracts the ratio of the maximum temperature difference to time during constant current discharge (F5) and the area covered by the temperature curve during discharge (F6) as temperature-related features; Incremental capacity related features: Incremental capacity analysis is a powerful non-invasive electrochemical analysis method that can explore subtle changes in electrochemical processes caused by capacity loss; the IC curve is the change curve of IC with discharge voltage, and the calculation process is as follows: Since the collected data has certain noise, it is necessary to pre-process the data. The IC curve is filtered by the SG filtering method. The position of the peak of the IC curve advances with the increase of the number of cycles, and the voltage value corresponding to the peak of the IC curve decreases with the increase of the number of cycles. Therefore, the position of the peak of the IC curve and the voltage value corresponding to the peak of the IC curve are extracted as IC-related features, the time corresponding to the peak of the incremental capacity curve (F7), and the voltage corresponding to the peak of the incremental capacity curve (F8); Differential thermovoltammetry related features: Differential thermovoltammetry is a signal analysis technique that can mine information related to battery degradation through changes in entropy; it determines the state of the battery by calculating the difference between temperature and voltage; DTV is calculated as follows: Where dT represents the differential value of temperature, and dV represents the differential value of voltage; The DTV curve still has a lot of noise. Similarly, the SG filtering method can be used to filter the original data. The position of the second trough of the DTV curve (F9) and the second trough value of the DTV curve (F10) are extracted as DTV related features. Energy-related characteristics: For different numbers of cycles, the time required for the battery to complete the constant current charging process is different, and the time required to complete the constant current discharging process is also different, resulting in different amounts of energy absorbed or released in each cycle; The energy-based method can calculate the constant current charging or discharging phase energy according to formula (12) to characterize the battery aging process. Where, respectively, are the start time and end time of charging or discharging, and respectively are the voltage and current at cycle i; Therefore, the energy (F11) in the constant current discharge stage can be extracted as an energy-related feature; Sample entropy related features: Sample entropy is a statistic used to measure the complexity of time series and plays an important role in analyzing signal complexity and predictability. As the battery degrades, the measurable voltage response under the same excitation will also change accordingly. The remarkable feature of this method is that at the end of each discharge cycle, the state of health (SOH) is estimated by simply calculating the sample entropy of the voltage, without involving complex calculation processes and avoiding the cumulative error caused by integration. Based on the above advantages, the sample entropy feature is widely used to describe the battery aging process. Given a time series \(X = \{x(1), x(2), \cdots, x(N)\}\) of length \(N\), the calculation method of sample entropy is as follows: Step1: Construct the time series \(X\) into an \(m\)-dimensional vector, that is \(X(i)=\{x(i),x(i + 1),\cdots,x(i + m-1)\}\ (13)\) where \(i = 1,2,\cdots,N - m+1\). \(X(i)\) represents a subsequence or a vector constructed from the original time series \(X\). Specifically, \(X(i)\) is a subsequence containing \(m\) consecutive data points, starting from the \(i\)-th data point of the original time series \(X\). Step2: Define the distance between vectors as \(i\neq j\), which is the maximum of the differences between their corresponding elements, that is: where \(d[X(i),X(j)]\) is the maximum distance between the corresponding elements of the two subsequences \(X(i)\) and \(X(j)\). \(x(i + k)\) is the \((k + 1)\)-th element in the subsequence \(X(i)\), and \(x(j + k)\) is the \((k + 1)\)-th element in the subsequence \(X(j)\), because the index usually starts from 0 or 1. Step3: Given a threshold \(r(r>0)\), count the number of \(d[X(i),X(j)]<r\) and calculate its ratio to the total number of vectors \(N - m\), that is In the formula, is a ratio value, which represents the ratio of the number of m-dimensional vectors similar to the i-th m-dimensional vector X(i) (the distance is less than r) to the total number of m-dimensional vectors in the time series. Here r is a preset threshold used to define the "similarity" between two vectors. N is the length of the original time series X, that is, the number of data points in the time series; m is the embedding dimension or pattern length, which determines the length of each subsequence (or vector) constructed from the original time series. Step4: Take the average of all the results obtained from Equation (8), that is: In the formula, B m (r) is an average ratio value, which represents the average ratio of vectors that are similar to at least one other m-dimensional vector (ie, the distance between them is less than a preset threshold r) among all m-dimensional vectors. Step5: Then increase the dimension \(m\) by 1 and repeat Steps 1 - 4; Step6: Then the theoretical sample entropy of this sequence is: However, in practice, \(N\) cannot be infinite but is a finite value, so the estimated value of the sample entropy is: Extract this sample entropy (F12) as a feature.

Citation Information

Patent Citations

  • Improved quantum particle swarm optimizing algorithm and method applied to prediction of network traffic

    CN107995027A

  • Fuel cell life prediction method

    CN111310387A

  • Lithium ion battery remaining service life prediction method and system based on IGS-SVM

    CN113281671A

  • Method for constructing short-term load prediction model of electric vehicle charging station

    CN116596129A

  • Method for predicting remaining service life of lithium-ion battery employing WDE-optimized LSTM network

    WO2020191800A1

Cited By

  • Lithium battery health state prediction method and device

    CN119619893A

  • Battery health state prediction method and device, equipment, storage medium and product

    CN119689272A

  • Battery health state estimation method based on end-cloud cooperation strategy

    CN119916222A

  • Semi-supervised lithium battery health state prediction method

    CN120085213A

  • Data driving fusion quantum optimization battery health state estimation method and system

    CN120142942A