High-precision estimation method for soc of power battery based on fusion lasso feature screening

Abstract

The application discloses a kind of power battery SOC high-precision estimation methods of fusion LASSO feature screening, comprising: collecting power battery multi-source working condition data and pre-processing, candidate feature pool is constructed;Sparse feature screening is carried out to feature variable in candidate feature pool using LASSO regression, and key feature subset with strong correlation with SOC is extracted;Based on the key feature subset obtained by screening, a data-driven SOC estimation model is established;Real-time acquisition of battery working condition data, input the trained SOC estimation model, and output SOC estimation value;The most representative feature subset is automatically screened out from a large number of candidate features by LASSO method, which effectively reduces the feature dimension and model complexity, while improving the generalization ability and estimation accuracy of the SOC estimation model, suitable for electric vehicle, energy storage system and other power battery management fields.

Description

Technical Field

[0001] This invention relates to the field of power battery technology, and in particular to a high-precision estimation method for power battery SOC that integrates LASSO feature screening. Background Technology

[0002] As the core energy storage unit in electric vehicles and energy storage systems, the accurate estimation of the state of charge (SOC) of the power battery is crucial for ensuring safe battery operation, optimizing energy management strategies, and extending battery life. SOC is analogous to the fuel gauge in a traditional gasoline vehicle, directly affecting the driver's range assessment and the battery's charge / discharge control strategy. However, SOC cannot be directly measured by sensors and must be indirectly estimated based on measurable parameters such as battery terminal voltage, current, and temperature. This makes SOC estimation one of the most challenging technical problems in battery management systems.

[0003] Currently, mainstream power battery SOC estimation methods are mainly divided into three categories: ampere-hour integration method, model method, and data-driven method. The ampere-hour integration method calculates SOC by integrating the current over time. This method is simple and easy to implement, but it suffers from initial SOC error accumulation and current measurement drift, leading to a significant increase in estimation error over long-term operation. The model method establishes the battery's state-space equation based on an equivalent circuit model and combines estimation algorithms such as Kalman filtering to achieve recursive SOC estimation. This method has high accuracy, but it is highly dependent on the accuracy of model parameters. Under battery aging and temperature changes, model parameters drift, leading to a decrease in estimation accuracy. The data-driven method directly learns the nonlinear mapping relationship between input features and SOC from historical data using machine learning algorithms. It has the advantage of not relying on a precise physical model. However, traditional feature engineering methods often rely on manual experience to select features, easily introducing redundant features or omitting key features, resulting in high model complexity and poor generalization ability.

[0004] To address the aforementioned issues, how to automatically select a subset of key features that are strongly correlated with SOC and are not redundant from a large number of candidate features, while simultaneously establishing a high-precision, low-complexity SOC estimation model, has become a pressing technical challenge in this field. Summary of the Invention

[0005] The purpose of this invention is to provide a high-precision SOC estimation method for power batteries that integrates LASSO feature selection, in order to solve the technical problems of existing SOC estimation methods, such as feature selection relying on human experience, high feature redundancy, and insufficient model generalization ability, thereby improving the accuracy and robustness of SOC estimation.

[0006] To achieve the above objectives, the technical solution adopted by this invention is: a high-precision estimation method for the state of charge (SOC) of a power battery that integrates LASSO feature screening, comprising the following steps:

[0007] S1. Collect multi-source operating condition data of the power battery and preprocess it to construct a candidate feature pool;

[0008] S2. Use LASSO regression to perform sparse feature screening on the feature variables in the candidate feature pool and extract the key feature subset that is strongly correlated with SOC.

[0009] S3. Based on the key feature subset obtained through screening, establish a data-driven SOC estimation model and train the model.

[0010] S4. Collect battery operating condition data in real time, input the collected data into the trained SOC estimation model, and output the SOC estimation value.

[0011] As a further improvement of the present invention, in S1, the multi-source operating condition data includes battery terminal voltage, charging and discharging current, battery surface temperature, battery internal temperature, cumulative charging and discharging capacity, number of charging and discharging cycles, battery open circuit voltage, and battery internal resistance.

[0012] As a further improvement of the present invention, in S1, the preprocessing includes data cleaning, missing value imputation, outlier removal and data normalization; the data normalization adopts the min-max normalization method to map each feature value to the [0,1] interval, thereby eliminating the influence of dimensional differences on feature selection and model training.

[0013] As a further improvement of the present invention, in S1, the candidate feature pool is constructed as follows:

[0014] The original operating condition data is subjected to time-domain feature extraction, frequency-domain feature extraction, and statistical feature extraction. The time-domain features include the rate of change of current, the rate of change of voltage, and the rate of change of temperature, which are used to capture the dynamic response characteristics of the battery. The frequency-domain features include the energy components of each frequency band obtained by fast Fourier transform, which are used to reflect the response characteristics of the battery under different frequency excitations. The statistical features include the mean, variance, maximum value, minimum value, skewness, and kurtosis, which are used to characterize the statistical regularity of the data distribution.

[0015] As a further improvement of the present invention, S2 specifically includes the following steps:

[0016] S21. Construct the objective function for the LASSO regression model:

[0017] ;

[0018] in, The total number of samples, For the first The true SOC value of each sample. For the first Candidate feature vectors of each sample Let be the dimension of the candidate features. For the regression coefficient vector, For regularization parameters, for Norm; the first term of the objective function is the mean squared error term, which measures the model's accuracy in fitting the SOC; the second term is... The regularization term imposes a sparsity constraint on the regression coefficients;

[0019] S22. Select the optimal regularization parameter λ through cross-validation: using... The cross-validation method randomly divides the training data into two parts. Each subset, taking turns... One subset is used as the training set and the remaining one subset is used as the validation set. A set of candidate λ values ​​is traversed, and the λ value that minimizes the mean square error of the validation set is selected as the optimal regularization parameter.

[0020] S23. Solve for the LASSO objective function to make the regression coefficient vector... The features corresponding to the non-zero coefficients are the key feature subsets obtained through screening.

[0021] As a further improvement of the present invention, in S3, the data-driven SOC estimation model is a support vector regression (SVR) model; the SVR model introduces... An insensitive loss function ensures good generalization performance while maintaining fitting accuracy. The radial basis function (RBF) kernel is chosen as the kernel function. ,in, The kernel width parameter controls the local response range of the kernel function. During model training, a selected subset of key features is used as the model input features, and the true SOC value is used as the model output target. The optimal penalty parameter is determined through grid search combined with cross-validation. and kernel parameters .

[0022] As a further improvement of the present invention, in S3, the data-driven SOC estimation model is a deep neural network model, including an input layer, at least two hidden layers, and an output layer, wherein the number of nodes in the input layer is equal to the number of features in the key feature subset. The hidden layer uses the ReLU activation function to introduce non-linearity, and the output layer has 1 node and uses a linear activation function to output the estimated SOC value. When training the model, the selected key feature subset is used as the model input features, and the true SOC value is used as the model output target. The gradient descent method or its variant algorithm is used to iteratively optimize the network weights until the loss function converges.

[0023] As a further improvement to the present invention, the following steps are also included:

[0024] S5. The output SOC estimate is fused and corrected with the SOC estimate obtained by the ampere-hour integration method; the details are as follows:

[0025] The SOC estimate obtained by the ampere-hour integral method is used as the state prediction value, and the SOC estimate output by S4 is used as the observation correction value. The Kalman filter algorithm is used for fusion estimation, and the final fused SOC estimate result is output.

[0026] The beneficial effects of this invention are:

[0027] 1. This invention introduces LASSO regression for feature selection, which can automatically identify the key feature subset most relevant to SOC from a large number of candidate features. It abandons the traditional method of relying on manual experience to select features, significantly reduces feature dimensionality and model complexity, and effectively avoids the "curse of dimensionality" problem.

[0028] 2. LASSO Regularized sparsity constraints can compress the coefficients of redundant features to zero. The selected feature subsets not only have strong predictive power, but also effectively suppress collinearity among features, thereby improving the generalization ability and robustness of the SOC estimation model.

[0029] 3. This invention uses a data-driven model for SOC estimation, which does not rely on an accurate battery physical model and a complex parameter identification process. This avoids the problem of model parameters drifting with battery aging and changes in operating conditions, and has strong adaptability to operating conditions.

[0030] 4. This invention further proposes a SOC estimation scheme that integrates LASSO screening and Kalman filtering. It combines the high-precision correction capability of the data-driven method with the continuous recursion property of the ampere-hour integration method, achieving high accuracy and strong robustness of SOC estimation under dynamic operating conditions, with the estimation error controllable within 1%. Attached Figure Description

[0031] Figure 1 This is an overall flowchart of an embodiment of the present invention;

[0032] Figure 2 This is a detailed flowchart of the LASSO feature selection process in Embodiment 2 of the present invention;

[0033] Figure 3 This is a schematic diagram of the candidate feature pool construction in Embodiment 2 of the present invention;

[0034] Figure 4 This is a flowchart of the SOC estimation method in Embodiment 3 of the present invention, which incorporates Kalman filtering. Detailed Implementation

[0035] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0036] Example 1

[0037] like Figure 1 As shown, a high-precision estimation method for the state of charge (SOC) of a power battery based on LASSO feature screening includes the following steps:

[0038] S1. Collect multi-source operating condition data of the power battery and preprocess it to construct a candidate feature pool;

[0039] S2. Use LASSO regression to perform sparse feature screening on the feature variables in the candidate feature pool and extract the key feature subset that is strongly correlated with SOC.

[0040] S3. Based on the key feature subset obtained from S2, establish a data-driven SOC estimation model and train the model.

[0041] S4. Collect battery operating condition data in real time, input the collected data into the trained SOC estimation model, and output the SOC estimation value.

[0042] Furthermore, the multi-source operating condition data mentioned in S1 includes battery terminal voltage, charge / discharge current, battery surface temperature, battery internal temperature, cumulative charge / discharge capacity, charge / discharge cycle count, battery open-circuit voltage, and battery internal resistance. These parameters can reflect the battery's operating state from multiple dimensions such as electrical characteristics, thermal characteristics, and aging characteristics, providing a rich data foundation for subsequent feature extraction.

[0043] Furthermore, the preprocessing described in S1 includes data cleaning, missing value imputation, outlier removal, and data normalization. Data normalization employs the min-max normalization method, mapping each feature value to the [0,1] interval to eliminate the impact of dimensional differences on feature selection and model training.

[0044] Furthermore, the specific methods for constructing the candidate feature pool in S1 include: extracting time-domain features, frequency-domain features, and statistical features from the original operating condition data. Time-domain features include current change rate, voltage change rate, and temperature change rate, used to capture the dynamic response characteristics of the battery; frequency-domain features include energy components of each frequency band obtained through Fast Fourier Transform, used to reflect the battery's response characteristics under different frequency excitations; statistical features include mean, variance, maximum value, minimum value, skewness, and kurtosis, used to characterize the statistical regularity of the data distribution. Through the above multi-dimensional feature extraction, an initial feature pool containing dozens to hundreds of candidate features is constructed.

[0045] Furthermore, the steps for sparse feature selection using LASSO regression in S2 are as follows:

[0046] S21. Construct the objective function for the LASSO regression model:

[0047]

[0048] in, The total number of samples, For the first The true SOC value of each sample. For the first Candidate feature vectors of each sample Let be the dimension of the candidate features. For the regression coefficient vector, For regularization parameters, for Norm. The first term of the objective function is the mean squared error term, which measures the model's accuracy in fitting the SOC; the second term is... The regularization term imposes a sparse constraint on the regression coefficients.

[0049] S22. Select the optimal regularization parameter λ through cross-validation. The cross-validation method randomly divides the training data into two parts. A subset (usually) Take 5 or 10), and alternately select one of them. One subset is used as the training set, and the remaining one subset is used as the validation set. A set of candidate λ values ​​is traversed, and the λ value that minimizes the mean square error of the validation set is selected as the optimal regularization parameter.

[0050] S23. Solve for the LASSO objective function to make the regression coefficient vector... The features corresponding to the non-zero coefficients are the key feature subsets obtained through screening. (LASSO's...) Regularization has automatic feature selection capabilities. During the optimization process, it can compress unimportant feature coefficients to zero, retaining only non-zero feature coefficients that are strongly correlated with SOC prediction, thereby significantly reducing feature dimensionality. LASSO's feature selection is a continuous shrinkage process that can simultaneously achieve variable shrinkage and automatic selection, overcoming the instability problem of traditional stepwise regression methods.

[0051] Furthermore, the data-driven SOC estimation model described in S3 is a Support Vector Regression (SVR) model. The SVR model introduces... The insensitive loss function ensures good generalization performance while maintaining fitting accuracy. Its kernel function is the radial basis function (RBF) kernel.

[0052]

[0053] in, The kernel width parameter controls the local response range of the kernel function.

[0054] As an alternative implementation, the data-driven SOC estimation model described in S3 is a deep neural network model. This model includes an input layer, at least two hidden layers, and an output layer. The number of nodes in the input layer is equal to the number of features in the key feature subset. The hidden layer uses the ReLU activation function to introduce nonlinearity, while the output layer has only one node and uses a linear activation function to output the SOC estimate. Deep neural networks can learn the highly nonlinear mapping relationship between input features and SOC, making them suitable for battery systems with complex dynamic characteristics.

[0055] Furthermore, in S3, when training the model, the selected subset of key features is used as the model input features, and the true SOC value is used as the model output target. For the SVR model, the optimal penalty parameters are determined through grid search combined with cross-validation. and kernel parameters For deep neural networks, gradient descent or its variants (such as the Adam optimizer) are used to iteratively optimize the network weights until the loss function converges.

[0056] Furthermore, S5 follows S4: the SOC estimate output from S4 is fused and corrected with the SOC estimate obtained by the ampere-hour integration method. The specific fusion method includes: using the SOC estimate obtained by the ampere-hour integration method as the state prediction value, and using the SOC estimate output from S4 as the observation correction value, a Kalman filter algorithm is used for fusion estimation, and the final fused SOC estimate is output. This method leverages the high time resolution of the ampere-hour integration method and the high-precision correction capability of the data-driven method, achieving complementary advantages and further improving the dynamic response performance and steady-state accuracy of SOC estimation.

[0057] Example 2

[0058] A high-precision method for estimating the state of charge (SOC) of a power battery by incorporating LASSO feature screening, the method comprising the following steps:

[0059] Step 1: Data Acquisition and Preprocessing

[0060] First, a power battery testing platform was built, using a specific model of ternary lithium-ion power battery (nominal voltage 3.7V, rated capacity 50Ah) as the test object. The ambient temperature was set to 25℃ in a constant temperature chamber, and various typical charging and discharging tests were conducted, including constant current charging and discharging, dynamic stress testing (DST), and the Chinese urban driving cycle (CLTC). The following multi-source operating condition data were simultaneously collected using a high-precision data acquisition system: battery terminal voltage (sampling frequency 1Hz), charging and discharging current (sampling frequency 1Hz), battery surface temperature (measured via thermocouple), cumulative charging and discharging capacity (obtained by integrating the current), number of charge and discharge cycles, and battery open-circuit voltage (measured after a sufficiently long period of rest).

[0061] After data acquisition, the raw data is preprocessed. First, data cleaning is performed to remove obviously erroneous data points caused by sensor malfunctions or communication anomalies. Then, linear interpolation is used to fill in missing values. For outliers exceeding reasonable physical ranges (such as voltages exceeding 4.2V or falling below 2.5V), [further processing is required]. The criteria are used to identify and eliminate features; finally, all feature variables are subjected to min-max normalization to map each feature value to the [0,1] interval, as shown in the formula:

[0062]

[0063] Where x is the original feature value. and These are the minimum and maximum values ​​of the feature in the training set, respectively.

[0064] Step 2: Construct a candidate feature pool:

[0065] like Figure 3 As shown, multi-dimensional feature extraction is performed on the preprocessed operating condition data to construct a candidate feature pool. Specifically, this includes:

[0066] (1) Time-domain feature extraction: Calculate the rate of change of current between adjacent sampling points. Voltage change rate and rate of temperature change This is used to capture the dynamic response characteristics of the battery. In addition, the sequence values ​​of current, voltage and temperature within a time window (e.g., 10 seconds, 30 seconds, 60 seconds) prior to the current moment are used as the original time-domain features.

[0067] (2) Frequency domain feature extraction: Fast Fourier transform is performed on the voltage and current signals within a fixed time window to extract the energy components of each frequency band (low frequency band 0-0.1Hz, mid frequency band 0.1-1Hz, high frequency band 1-5Hz) as frequency domain features to reflect the impedance characteristics of the battery under different frequency excitations.

[0068] (3) Statistical feature extraction: Calculate the statistics of each physical quantity within a fixed time window, including mean, variance, maximum value, minimum value, range, skewness (measures the asymmetry of distribution) and kurtosis (measures the sharpness of distribution).

[0069] Through the above multidimensional feature extraction, approximately 150 candidate features were obtained, forming the initial candidate feature pool; ,in For the sample size, Let be the feature dimension.

[0070] Step 3: LASSO Feature Filtering:

[0071] like Figure 2 As shown, LASSO regression is used to filter sparse features in the candidate feature pool. The specific process of this step is as follows:

[0072] Step 3.1: Construct the objective function for LASSO regression:

[0073]

[0074] in, For the first The true SOC value of each sample (reference value obtained through high-precision testing equipment). For the first Candidate feature vectors of each sample Let be the vector of regression coefficients to be solved, and λ be the regularization parameter. This objective function minimizes the SOC fitting error while simultaneously... Norm constraints force some regression coefficients to shrink to zero, thus enabling feature selection.

[0075] Step 3.2: Select the optimal regularization parameter λ using 10-fold cross-validation. Randomly divide the training data into 10 subsets, and alternately use 9 subsets as the training set and the remaining subset as the validation set. Iterate through a set of candidate λ values ​​that are distributed in a geometric progression (e.g., from...). arrive (A total of 50 candidate values ​​are available). For each λ value, the mean squared error of 10 validations is calculated, and the λ with the smallest mean squared error is selected as the optimal regularization parameter. .

[0076] Step 3.3: In the selected Solving the LASSO objective function yields the regression coefficient vector. Filter out The feature variables corresponding to the non-zero coefficients constitute a subset of key features. In this embodiment, approximately 12-18 key features are selected from 150 candidate features, reducing the feature dimension by approximately 88%-92%. Typical selected features include: voltage change rate, temperature, current integral value, previous SOC estimate, and voltage variance within the time window.

[0077] Step 4: Build and train the SVR model:

[0078] Using the key feature subset obtained in step 3 as input features and the true SOC value as the output target, a Support Vector Regression (SVR) model is established. The radial basis function kernel function is selected for the SVR model.

[0079]

[0080] Hyperparameters that need to be optimized during model training include: penalty parameters (Control model for exceeding) The degree of penalty for the pipeline samples), kernel width parameter (Controlling the local response range of the radial basis functions) and Insensitive loss parameters (controlling fitting accuracy). The optimal hyperparameter combination is determined using a grid search combined with 5-fold cross-validation. The candidate value is , The candidate value is , The candidate value is The parameter combination that minimizes the mean squared error of cross-validation is selected. In this embodiment, the optimal parameters are: , , .

[0081] The training set data is input into the SVR model, and the model is trained using the selected hyperparameters. For the SVR model, the training process is equivalent to solving a convex quadratic programming problem to obtain the support vectors and their corresponding weight coefficients, thereby constructing the decision function.

[0082] Step 5: Online SOC estimation:

[0083] During the actual operation of the power battery, the battery's terminal voltage, current, temperature, and other operating condition data are collected in real time at the same sampling frequency. After the collected data undergoes the same preprocessing and feature extraction process, the key feature subset selected in step 3 is extracted from the candidate features and input into the trained SVR model. The model output is the SOC estimate at the current moment.

[0084] Example 3

[0085] This embodiment, based on embodiments 1 and 2, further introduces a fusion correction mechanism with the ampere-hour integration method to improve the robustness and dynamic response performance of SOC estimation under long-term operation and complex working conditions. For example... Figure 4 As shown, the method in this embodiment includes:

[0086] Steps 1-4 are the same as steps 1-4 in Example 2, obtaining the SOC estimate output by the LASSO-SVR model. .

[0087] Step 5: SOC estimation using the ampere-hour integration method:

[0088] The SOC estimate is calculated in real time using the ampere-hour integration method:

[0089]

[0090] in, The initial SOC estimate is given. For Coulomb efficiency, For the battery's rated capacity, Let be the discharge current (positive for discharge, negative for charging). The ampere-hour integration method can provide a high time resolution continuous estimate of SOC in a short time, but it suffers from error accumulation.

[0091] Step 6: Kalman filter fusion estimation:

[0092] Construct discrete state-space equations and use the SOC estimate from the ampere-hour integral method. As a state prediction value, the SOC estimate from the LASSO-SVR model is used. As an observation correction, the system state equation and observation equation are as follows:

[0093] Equations of state: ;

[0094] Observation equation: ;

[0095] in, Represents the discrete time step. The sampling interval is... For process noise, This is observation noise. Observation value Take the estimated output of the LASSO-SVR model State estimation is performed using a Kalman filter recursive formula:

[0096] (1) State prediction: ;

[0097] (2) Error covariance prediction: ;

[0098] (3) Kalman gain calculation: ;

[0099] (4) Status update: ;

[0100] (5) Error covariance update: ;

[0101] in, For process noise covariance, To observe the noise covariance. and The value is determined through empirical debugging: Take the smaller value (e.g.) Using the dynamic recursion of the trust ampere-hour integral, A relatively large value is chosen to balance the weights of observation corrections. The final SOC estimate after fusion is output. .

[0102] The fused SOC estimation method combines the high time resolution and continuous recursion capability of the ampere-hour integration method with the high accuracy correction capability of the LASSO-SVR model. Experimental results show that after running 10 complete cycles under dynamic stress testing conditions, the cumulative error of the fused method does not exceed 0.8%, which is significantly better than the estimation performance of the single method.

[0103] Example 4

[0104] The difference between this embodiment and embodiment 2 is that in step 4, a deep neural network (DNN) model is used instead of the SVR model as the data-driven SOC estimation model.

[0105] The deep neural network model is structured as follows: the number of nodes in the input layer is equal to the number of key features obtained by LASSO (approximately 15 in this embodiment); the first hidden layer contains 64 nodes and uses the ReLU activation function; the second hidden layer contains 32 nodes and uses the ReLU activation function; the third hidden layer contains 16 nodes and uses the ReLU activation function; the output layer contains 1 node and uses a linear activation function to output the SOC estimate.

[0106] The model was trained using the Adam optimizer with an initial learning rate of 0.001, a batch size of 32, and 200 training epochs. Mean squared error (MSE) was used as the loss function. To prevent overfitting, a Dropout layer was added after each hidden layer with a Dropout ratio of 0.2. Early stopping was employed, terminating training prematurely if the validation set loss did not improve within 20 consecutive epochs.

[0107] Using a small feature set filtered by LASSO as input to the DNN model significantly reduces the number of parameters and training difficulty while maintaining high estimation accuracy. Under the same test conditions, the DNN model's mean absolute error for SOC estimation is approximately 0.45%, slightly better than the SVR model.

[0108] The embodiments described above are merely illustrative of specific implementations of the present invention, and while the descriptions are detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention.

Claims

1. A high-precision estimation method for the state of charge (SOC) of a power battery integrating LASSO feature screening, characterized in that, Includes the following steps: S1. Collect multi-source operating condition data of the power battery and preprocess it to construct a candidate feature pool; S2. Use LASSO regression to perform sparse feature screening on the feature variables in the candidate feature pool and extract the key feature subset that is strongly correlated with SOC. S3. Based on the key feature subset obtained through screening, establish a data-driven SOC estimation model and train the model. S4. Collect battery operating condition data in real time, input the collected data into the trained SOC estimation model, and output the SOC estimation value.

2. The high-precision estimation method for power battery SOC based on LASSO feature screening according to claim 1, characterized in that, In S1, the multi-source operating condition data includes battery terminal voltage, charging and discharging current, battery surface temperature, battery internal temperature, cumulative charging and discharging capacity, number of charging and discharging cycles, battery open circuit voltage, and battery internal resistance.

3. The high-precision estimation method for power battery SOC based on LASSO feature screening according to claim 2, characterized in that, In S1, the preprocessing includes data cleaning, missing value imputation, outlier removal, and data normalization. The data normalization process uses the min-max normalization method to map each feature value to the [0,1] interval, eliminating the impact of dimensional differences on feature selection and model training.

4. The high-precision estimation method for power battery SOC based on LASSO feature screening according to claim 3, characterized in that, In S1, the candidate feature pool is constructed as follows: The original operating condition data is subjected to time-domain feature extraction, frequency-domain feature extraction, and statistical feature extraction. The time-domain features include the rate of change of current, the rate of change of voltage, and the rate of change of temperature, which are used to capture the dynamic response characteristics of the battery. The frequency-domain features include the energy components of each frequency band obtained by fast Fourier transform, which are used to reflect the response characteristics of the battery under different frequency excitations. Statistical characteristics include mean, variance, maximum, minimum, skewness, and kurtosis, which are used to characterize the statistical regularity of data distribution.

5. The high-precision estimation method for power battery SOC based on LASSO feature screening according to claim 1, characterized in that, S2 specifically includes the following steps: S21. Construct the objective function for the LASSO regression model: ； in, The total number of samples, For the first The true SOC value of each sample. For the first Candidate feature vectors of each sample Let be the dimension of the candidate features. For the regression coefficient vector, For regularization parameters, for Norm; the first term of the objective function is the mean squared error term, which measures the model's accuracy in fitting the SOC; the second term is... The regularization term imposes a sparsity constraint on the regression coefficients; S22. Select the optimal regularization parameter λ through cross-validation: using... The cross-validation method randomly divides the training data into two parts. Each subset, taking turns... One subset is used as the training set and the remaining one subset is used as the validation set. A set of candidate λ values ​​is traversed, and the λ value that minimizes the mean square error of the validation set is selected as the optimal regularization parameter. S23. Solve for the LASSO objective function to make the regression coefficient vector... The features corresponding to the non-zero coefficients are the key feature subsets obtained through screening.

6. The high-precision estimation method for power battery SOC based on LASSO feature screening according to claim 1, characterized in that, In S3, the data-driven SOC estimation model is a support vector regression (SVR) model; the SVR model introduces... An insensitive loss function ensures good generalization performance while maintaining fitting accuracy. The radial basis function (RBF) kernel is chosen as the kernel function. ,in, The kernel width parameter controls the local response range of the kernel function. During model training, a selected subset of key features is used as the model input features, and the true SOC value is used as the model output target. The optimal penalty parameter is determined through grid search combined with cross-validation. and kernel parameters .

7. The high-precision estimation method for power battery SOC based on LASSO feature screening according to claim 1, characterized in that, In S3, the data-driven SOC estimation model is a deep neural network model, including an input layer, at least two hidden layers, and an output layer. The number of nodes in the input layer is equal to the number of features in the key feature subset. The hidden layer uses the ReLU activation function to introduce nonlinearity, and the output layer has 1 node and uses a linear activation function to output the SOC estimate. When training the model, the selected subset of key features is used as the input features of the model, the true SOC value is used as the output target of the model, and the gradient descent method or its variant algorithm is used to iteratively optimize the network weights until the loss function converges.

8. The high-precision estimation method for power battery SOC based on LASSO feature screening according to any one of claims 1-7, characterized in that, It also includes the following steps: S5. The output SOC estimate is fused and corrected with the SOC estimate obtained by the ampere-hour integration method; the details are as follows: The SOC estimate obtained by the ampere-hour integral method is used as the state prediction value, and the SOC estimate output by S4 is used as the observation correction value. The Kalman filter algorithm is used for fusion estimation, and the final fused SOC estimate result is output.

Images