A battery SOC estimation method based on NYSR-SVMR model

Abstract

This invention discloses a battery SOC estimation method based on the NYSR-SVMR model, comprising the following steps: S1: collecting battery voltage, current, and SOC information under various operating conditions; S2: selecting a training set D = (x i ,y i ), i = 1, 2, ..., n, where x i Let y be the set of measured values ​​of current, voltage, and temperature for the i-th sample battery. i Let S1 be the battery SOC value of the i-th sample, and n represent the number of samples; S3: Normalize the selected data; S4: Initialize the model and train it; S5: Validate the model. This can effectively improve the estimation performance and accuracy of battery SOC.

Description

Technical Field

[0001] This invention relates to the field of lithium battery computing technology, specifically to a battery SOC estimation method based on the NYSR-SVMR model. Background Technology

[0002] The State of Charge (SOC) of a battery refers to the state of its remaining charge. The accuracy of calculating this remaining charge has always been a challenge in the industry. In particular, as batteries degrade with use, their capacity decreases, leading to increasingly larger errors in the calculated SOC value.

[0003] State of Charge (SOC) estimation methods can be broadly categorized into three types: parametric measurement methods, model-based methods, and data-driven methods. Parametric measurement methods include open-circuit voltage methods, coulomb counting methods, and impedance measurement methods. While simple to implement, these methods suffer from low practicality and poor estimation accuracy. Model-based methods rely on an understanding and mathematical modeling of the physical and chemical processes of the battery system, including Kalman filtering, particle filtering, and sliding mode observer methods. These methods offer high estimation accuracy, but model building is complex, and accuracy is highly dependent on model accuracy. Data-driven methods include neural network methods and support vector machine methods. While these methods provide accurate estimation results, they are computationally intensive and require high processing power from the chip. Therefore, there is an urgent need for an estimation method that can improve the performance of battery SOC estimation. Summary of the Invention

[0004] This invention aims to solve the technical problems existing in the prior art, and innovatively proposes a battery SOC estimation method based on the NYSR-SVMR model, which can effectively improve the estimation performance of battery SOC and increase the estimation accuracy of battery SOC.

[0005] To achieve the above objectives, this invention provides a battery SOC estimation method based on the NYSR-SVMR model, comprising the following steps:

[0006] S1: Collect battery voltage, current, and SOC information under various operating conditions;

[0007] S2: Select the training set from the collected information. ,in For the first A collection of measured values ​​of current, voltage, and temperature for each sample battery. For the first The battery SOC value of each sample Indicates the number of samples;

[0008] S3: Normalize the selected data;

[0009] S4: Initialize the model and train it;

[0010] S5: Validate the model.

[0011] In the above scheme, step S4 also includes the following steps:

[0012] S4-1: The model is set as follows:

[0013]

[0014] in, This is an estimate of the battery's SOC. It is a collection of measured values ​​for the battery's current, voltage, and temperature. For coefficients, This is the deviation value;

[0015] S4-2: Accelerate the training speed of the support vector machine model by using the Nystrom approximation algorithm to obtain NYSR-SVMR;

[0016] S4-3: Select the training set Substitute the values ​​into the defined model and train the support vector machine model.

[0017] In the above scheme, step S4-3 also includes the following steps:

[0018] S4-3-1: Calculated using the following formula:

[0019]

[0020] in, Let be the objective function. To solve Minimum Parameter , and For the introduced slack variables, The penalty coefficient is... For approximate accuracy, and with the following constraints:

[0021]

[0022] in, For training set The Middle The SOC value of the battery for each sample;

[0023] S4-3-2: Simplify the formula in step S4-3-1;

[0024] The loss function is simplified as follows using equality constraints and the L2 norm:

[0025]

[0026] in, For training set The Middle A collection of measurements of current, voltage, and temperature of a sample battery; To sample Mapped to The mapping function, i.e. ; For error variables, For regularization parameters;

[0027] S4-3-3: Construct the Lagrange function as follows:

[0028]

[0029] in, For Lagrange multipliers, For the sample The mapping function, For the objective function, For training set The Middle The SOC value of the battery for each sample;

[0030] S4-3-4: Optimize the Lagrange function according to the optimization conditions;

[0031] S4-3-5: Suppose the kernel function is as follows:

[0032]

[0033] For kernel function, for The mapping function, For training set The Middle A collection of measurements of current, voltage, and temperature of a sample battery;

[0034] S4-3-6: Substitute the function from step S4-3-5 into the optimized Lagrangian function from step S4-3-4 for calculation, and obtain the following equation:

[0035]

[0036] in, It is the identity matrix. ;

[0037] , The kernel matrix is ​​composed of kernel functions, and its constituent elements are... ,in ;

[0038] S4-3-7: Solve the equation in S4-3-6 to obtain the parameters. and The prediction model in the dual space is constructed as follows:

[0039] .

[0040] In the above scheme, step S4-2 also includes the following steps:

[0041] S4-2-1: According to Construct the kernel matrix using the number of samples; due to If the number of samples is n, then the kernel matrix... The scale is ;

[0042] S4-2-2: From the kernel matrix If M samples are drawn from the sample, then the kernel matrix of the M samples is... The scale is ;

[0043] Then the kernel matrix The eigenvector matrix is ​​as follows: ;

[0044] in The eigenvector matrix, , A diagonal matrix composed of eigenvalues. ;

[0045] Then the kernel matrix The eigenvector matrix is ​​as follows: ;

[0046] in The eigenvector matrix, , It is a diagonal matrix composed of eigenvalues. ;

[0047] S4-2-3: Decompose the vector;

[0048] S4-2-4: Calculated using the following formula:

[0049]

[0050] in For the kernel matrix The Middle Line 1 Column elements, , and These are the eigenvalues ​​and their corresponding eigenfunctions;

[0051] S4-2-5: Transform the formula in step S4-2-4 using integral transformation. The transformed formula is as follows:

[0052]

[0053] in, Input variables The probability density function, and These are the eigenvalues ​​and their corresponding eigenfunctions;

[0054] S4-2-6: Draw M samples from N samples. and set samples The characteristic functions are as follows:

[0055]

[0056] S4-2-7: Combine the vectors decomposed in step S4-2-3 with the function in step S4-2-6 to obtain a high-dimensional vector. The i-th element and the i-th The relational expressions for the characteristic functions are as follows:

[0057]

[0058] Where M is the number of samples drawn. Eigenvalues The corresponding estimated value, For matrix The Middle Line 1 Column elements, for No. Line 1 eigenvectors of the column;

[0059] And obtain The estimated values ​​are as follows: ;

[0060] S4-2-8: The result obtained in step S4-2-7 The estimated values ​​are substituted into the prediction model constructed in S4-3-7, and the prediction model is rewritten as follows:

[0061]

[0062] The parameters of the original spatial model are obtained using the following formula. and b:

[0063] .

[0064] In the above scheme, step S4-3-4 includes the following steps:

[0065] The Karush-Kuhn-Tucker conditions are as follows:

[0066]

[0067] Based on the Karush-Kuhn-Tucker conditions, we can obtain:

[0068] .

[0069] In the above scheme, step S4-2-3 also includes the following steps:

[0070] The expression is as follows:

[0071]

[0072] The expression is as follows:

[0073]

[0074] The expression after vector decomposition is as follows:

[0075]

[0076] in, and These are the eigenvalues ​​and their corresponding eigenfunctions. Characteristic function The corresponding estimated value, Eigenvalues The corresponding estimated value.

[0077] In the above scheme, step S5 includes the following steps:

[0078] S5-1: Group the data collected in step S1 into training set D and test set X, where test set X consists of the measured values ​​of battery voltage and current.

[0079] S5-2: Test the model trained in step S4 based on the test set X.

[0080] In the above scheme, step S5-2 also includes the following steps:

[0081] S5-2-1: Calculate the root mean square error value using the following formula:

[0082]

[0083] in, denoted as the root mean square error value. For data length, For training set The Middle The battery SOC value of each sample This is the estimated SOC value;

[0084] S5-2-2: Calculate the maximum absolute error value using the following formula:

[0085]

[0086] in, The value of the maximum absolute error is . For data length, For training set The Middle The battery SOC value of each sample This is the estimated SOC value.

[0087] In summary, the beneficial effects of this invention are: it enables direct solution of the original SVMR model using the Nystrom algorithm, improving computational efficiency, reducing training time, saving computational resources, and effectively improving the estimation performance and accuracy of battery SOC. The established training and test sets are used to test the learned model, thereby verifying the modeling ability of the learning machine and ensuring prediction performance and estimation accuracy. Attached Figure Description

[0088] Figure 1 This is a graph showing the current variation of an INR18650 lithium battery under DST conditions.

[0089] Figure 2 This is a graph showing the voltage variation of an INR18650 lithium battery under DST conditions.

[0090] Figure 3 This is a graph showing the SOC change of the INR18650 lithium battery under DST conditions.

[0091] Figure 4 This is the SOC estimate of the INR18650 lithium battery based on the LS-SVMR method.

[0092] Figure 5 This is the estimated variance of the INR18650 lithium battery based on the LS-SVMR method.

[0093] Figure 6This is the estimated SOC value of the INR18650 lithium battery under the estimation method of this technical solution.

[0094] Figure 7 This is the estimated variance of the estimation method for the INR18650 lithium battery in this technical solution. Detailed Implementation

[0095] The present invention will be further described below with reference to the embodiments and accompanying drawings:

[0096] This invention provides a battery SOC estimation method based on the NYSR-SVMR model, comprising the following steps:

[0097] S1: Collect information on battery voltage, current, and estimated battery SOC under various operating conditions;

[0098] S2: Select the training set from the collected information. ,in For the first A collection of measured values ​​of current, voltage, and temperature for each sample battery. For the first The battery SOC value of each sample Indicates the number of samples;

[0099] S3: Normalize the selected data.

[0100] S4: Initialize the model and train it;

[0101] S4-1: The model is set as follows:

[0102]

[0103] Among them, among them, This is an estimate of the battery's SOC. It is a collection of measured values ​​for the battery's current, voltage, and temperature. For coefficients, This is the deviation value;

[0104] S4-2: Calculation of the original space of the LS-SVMR model using NYSR-SVMR ;

[0105] S4-2-1: According to Construct the kernel matrix using the number of samples; due to If the number of samples is n, then the kernel matrix... The scale is , where n=N;

[0106] S4-2-2: From the kernel matrix If M samples are drawn from the sample, then the kernel matrix of the M samples is... The scale is ;

[0107] Then the kernel matrix The eigenvector matrix is ​​as follows: ;

[0108] in The eigenvector matrix, , It is a diagonal matrix composed of eigenvalues. ;

[0109] Then the kernel matrix The eigenvector matrix is ​​as follows: ;

[0110] in The eigenvector matrix, , It is a diagonal matrix composed of eigenvalues. ;

[0111] S4-2-3: Decompose the vector;

[0112] The expression is as follows:

[0113]

[0114] The expression is as follows:

[0115]

[0116] right The decomposed vectors are as follows:

[0117]

[0118] in, and These are the eigenvalues ​​and their corresponding eigenfunctions. Characteristic function The corresponding estimated value, Eigenvalues The corresponding estimated value;

[0119] S4-2-4: Calculated using the following formula:

[0120]

[0121] in For the kernel matrix The Line 1 Column elements, , and These are the eigenvalues ​​and their corresponding eigenfunctions;

[0122] S4-2-5: Transform the formula in step S4-3-4 using integral transformation. The transformed formula is as follows:

[0123]

[0124] in, Input variables The probability density function, and These are the eigenvalues ​​and their corresponding eigenfunctions;

[0125] S4-2-6: Draw M samples from N samples. and set samples The characteristic functions are as follows:

[0126]

[0127] S4-2-7: Combine the vectors decomposed in step S4-2-3 with the function in step S4-2-6 to obtain a high-dimensional vector. The The element and the first The relational expressions for the characteristic functions are as follows:

[0128]

[0129] Where M is the number of samples drawn. Eigenvalues The corresponding estimated value, For matrix The Line 1 Column elements, for No. Line 1 eigenvectors of the column;

[0130] And obtain The estimated values ​​are as follows: ;

[0131] S4-2-8: The result obtained in step S4-2-7 The estimated values ​​are substituted into the prediction model constructed in S4-2-7, and the prediction model is rewritten as follows:

[0132]

[0133] The parameters of the original spatial model are obtained using the following formula. and b:

[0134] ;

[0135] S4-3: Substitute the selected training set D into the set model, and train the model using a support vector machine;

[0136] S4-3-1: Calculated using the following formula:

[0137]

[0138] in, Let be the objective function. To solve Minimum Parameter , and For the introduced slack variables, The penalty coefficient is... For approximate accuracy, and with the following constraints:

[0139]

[0140] in, For training set The Middle The SOC value of the battery for each sample;

[0141] Slack variables are a technique used to solve linearly inseparable problems. In the case of linear inseparability, SVMR allows some data points to lie on the wrong side of the hyperplane and tolerates these errors by introducing slack variables. These slack variables allow data points to have a certain distance between the wrong side and the hyperplane, thereby improving the robustness of the model. c is the penalty coefficient, also known as the regularization parameter, which controls how well the model fits the training data and how much it penalizes complexity. A larger penalty coefficient leads to a more rigorous fit to the training data, potentially resulting in a simpler model and avoiding overfitting, but may lead to a larger training error. A smaller penalty coefficient allows the model to fit the training data more flexibly, potentially leading to a more complex model and overfitting, but with a smaller training error. Therefore, the choice of penalty coefficient needs to be adjusted according to the specific problem and dataset to achieve a balance between bias and variance for optimal generalization performance. Approximate accuracy is one of the metrics used to evaluate the performance of support vector machine models. It usually refers to the model's performance on data outside the training set. Higher approximate accuracy means that the model performs well on unseen data and has better generalization ability.

[0142] S4-3-2: The formula in step S4-2-1 is simplified as follows by using equality constraints and the L2 norm as the loss function:

[0143]

[0144] in, For training set The Middle A collection of measurements of current, voltage, and temperature from individual battery samples. To sample Mapped to The mapping function, i.e. ; For error variables, For regularization parameters;

[0145] S4-3-3: Construct the Lagrange function as follows:

[0146]

[0147] in, For Lagrange multipliers, For the sample The mapping function, For the objective function, For training set The Middle The SOC value of the battery for each sample;

[0148] S4-3-4: Optimize according to the optimization conditions;

[0149] The optimization conditions are as follows:

[0150]

[0151] The optimized Lagrange function is as follows:

[0152]

[0153] S4-3-5: Suppose the kernel function is as follows:

[0154]

[0155] For kernel function, for The mapping function, For training set The Middle A collection of measurements of current, voltage, and temperature of a sample battery;

[0156] S4-3-6: Substitute the function from step S4-3-5 into the optimized Lagrangian function from step S4-3-4 for calculation, and obtain the following equation:

[0157]

[0158] in, It is the identity matrix. ;

[0159] , The kernel matrix is ​​composed of kernel functions, and its constituent elements are... ,in ;

[0160] S4-3-7: Solve the equation in S4-3-6 to obtain the parameters. and The prediction model in the dual space is constructed as follows:

[0161]

[0162] S5: Validate the model;

[0163] S5-1: Group the data collected in step S1 into training set D and test set X, where test set X consists of the measured values ​​of battery voltage and current.

[0164] S5-2: Test the model trained in step S4 based on the test set X.

[0165] S5-2-1: Calculate the root mean square error value using the following formula:

[0166]

[0167] in, denoted as the root mean square error value. For data length, For training set The Middle The battery SOC value of each sample This is the estimated SOC value;

[0168] S5-2-2: Calculate the maximum absolute error value using the following formula:

[0169]

[0170] in, The value of the maximum absolute error is . For data length, For training set The Middle The battery SOC value of each sample This is the estimated SOC value.

[0171] To verify the effectiveness of the proposed method, the following uses charge-discharge data of an INR18650 lithium battery under dynamic stress testing (DST) conditions. The selected INR18650 lithium battery has a rated capacity of 2000mAh and a rated voltage of 3.7V. The selected battery charge-discharge data comes from open-source data from the University of Maryland. After normalization of the open-source data, 1062 sets were selected for testing at a sampling interval of 10s. The current change curve, voltage change curve, and battery SOC change curve in the selected data are shown below. Figures 1-3 As shown.

[0172] In addition, SOC estimation based on the LS-SVMR method was added as a comparative experiment to better verify the reliability and superiority of the proposed algorithm. The battery SOC estimates and estimation error obtained through the experiment are shown in the figure below. Figures 4-7 As shown.

[0173] Figure 4 The SOC curve is estimated using the LS-SVMR method. Figure 5 Estimate its error by Figure 4 , 5 It can be seen that when using the LS-SVMR method for SOC estimation, the estimated value can track the true value well, but there is a significant error.

[0174] By comparing the performance metrics of the battery SOC values ​​estimated using various models, a performance comparison of the battery SOC estimated by each model is shown in Table 1.

[0175] Table 1 Comparison of estimated battery SOC performance by different models

[0176] method RMSE / % MAE / % Maximum error / % The proposed method 0.01 0.007 0.12 LS-SVMR 1.16 0.9 2.30

[0177] pass Figures 4-7 As shown in Table 1, compared with the LS-SVMR method for estimating battery SOC, this technical solution has smaller RMSE, MAE and maximum error when estimating battery SOC. Therefore, this technical solution has better performance and can effectively improve the estimation performance of battery SOC and increase the estimation accuracy of battery SOC.

Claims

1. A battery SOC estimation method based on the NYSR-SVMR model, characterized in that: The following steps are involved: S1: Collect battery voltage, current, and SOC information under various operating conditions; S2: Select the training set from the collected information. ,in For the first A collection of measured values ​​of current, voltage, and temperature for each sample battery. For the first The battery SOC value of each sample Indicates the number of samples; S3: Normalize the selected data; S4: Initialize the model and train it; S4-1: The model is set as follows: ； in, This is an estimate of the battery's SOC. It is a collection of measured values ​​for the battery's current, voltage, and temperature. For coefficients, This is the deviation value; S4-2: Accelerate the training speed of the support vector machine model by using the Nystrom approximation algorithm to obtain NYSR-SVMR; S4-3: Select the training set Substitute the values ​​into the defined model and train the support vector machine model; S5: Validate the model; S5-1: Group the data collected in step S1 into training set D and test set X, where test set X consists of the measured values ​​of battery voltage and current. S5-2: Test the model trained in step S4 according to the test set X; S5-2-1: Calculate the root mean square error value using the following formula: ； in, is the root mean square error value. For data length, For training set The Middle The battery SOC value of each sample This is the estimated SOC value; S5-2-2: Calculate the maximum absolute error value using the following formula: ； in, The value of the maximum absolute error is . For data length, For training set The Middle The battery SOC value of each sample This is the estimated SOC value.

2. The battery SOC estimation method based on the NYSR-SVMR model according to claim 1, characterized in that: Step S4-3 also includes the following steps: S4-3-1: Calculated using the following formula: ； in, Let be the objective function. To solve Minimum Parameter , and For the introduced slack variables, The penalty coefficient is... For approximate accuracy, and with the following constraints: ； in, For training set The Middle The SOC value of the battery for each sample; S4-3-2: Simplify the formula in step S4-3-1; The loss function is simplified as follows using equality constraints and the L2 norm: ； in, For training set The Middle A collection of measurements of current, voltage, and temperature of a sample battery; To sample Mapped to The mapping function, i.e. ; For error variables, For regularization parameters; S4-3-3: Construct the Lagrange function as follows: ； in, For Lagrange multipliers, For the sample The mapping function, For the objective function, For training set The Middle The SOC value of the battery for each sample; S4-3-4: Optimize the Lagrange function according to the optimization conditions; S4-3-5: Suppose the kernel function is as follows: ； in, For kernel function, for The mapping function, For training set The Middle A collection of measurements of current, voltage, and temperature of a sample battery; S4-3-6: Substitute the function from step S4-3-5 into the optimized Lagrangian function from step S4-3-4 for calculation, and obtain the following equation: ； in, It is the identity matrix. ; , The kernel matrix is ​​composed of kernel functions, and its constituent elements are... ,in ; S4-3-7: Solve the equation in S4-3-6 to obtain the parameters. and The prediction model in the dual space is constructed as follows: 。 3. The battery SOC estimation method based on the NYSR-SVMR model according to claim 1, characterized in that: Step S4-2 also includes the following steps: S4-2-1: According to Construct the kernel matrix using the number of samples; due to If the number of samples is n, then the kernel matrix... The scale is ; S4-2-2: From the kernel matrix If M samples are drawn from the sample, then the kernel matrix of the M samples is... The scale is ; Then the kernel matrix The eigenvector matrix is ​​as follows: ; in The eigenvector matrix, , A diagonal matrix composed of eigenvalues. ; Then the kernel matrix The eigenvector matrix is ​​as follows: ; in The eigenvector matrix, , A diagonal matrix composed of eigenvalues. ; S4-2-3: Decompose the vector; S4-2-4: Calculated using the following formula: ； in For the kernel matrix The Middle Line number Column elements, , and These are the eigenvalues ​​and their corresponding eigenfunctions; S4-2-5: Transform the formula in step S4-3-4 using integral transformation. The transformed formula is as follows: ； in, Input variables The probability density function, and These are the eigenvalues ​​and their corresponding eigenfunctions; S4-2-6: Draw M samples from N samples. and set samples The characteristic functions are as follows: ； S4-2-7: Combine the vectors decomposed in step S4-2-3 with the function in step S4-2-6 to obtain a high-dimensional vector. The i-th element and the i-th The relational expressions for the characteristic functions are as follows: ； Where M is the number of samples drawn. Eigenvalues The corresponding estimated value, For matrix The Middle Line number Column elements, for No. Line number The eigenvectors of the column; And obtain The estimated values ​​are as follows: ; S4-2-8: The result obtained in step S4-2-7 The estimated values ​​are substituted into the prediction model constructed in S4-2-7, and the prediction model is rewritten as follows: ； The parameters of the original spatial model are obtained using the following formula. and b: 。 4. The battery SOC estimation method based on the NYSR-SVMR model according to claim 3, characterized in that: Step S4-2-4 includes the following steps: The conditions for Karush-Kuhn-Tucker are as follows: ； Based on the Karush-Kuhn-Tucker conditions, we can obtain: 。 5. The battery SOC estimation method based on the NYSR-SVMR model according to claim 3, characterized in that: Step S4-2-3 also includes the following steps: The expression is as follows: ； The expression is as follows: ； The expression after vector decomposition is as follows: ； in, and These are the eigenvalues ​​and their corresponding eigenfunctions. Characteristic function The corresponding estimated value, Eigenvalues The corresponding estimated value.

Images