EIS Prediction Method and Device for Lithium Batteries Used in Power Energy Storage

Abstract

The present invention provides a method for predicting the EIS of lithium batteries for power energy storage, belonging to the field of impedance spectrum prediction of lithium batteries. It includes dividing the preprocessed battery sample data into a test set, a data set one, and a data set two, and calculating the geometric parameters of the battery in different states according to the impedance spectrum; using the data set one as the input and the geometric parameters as the output to train the MLP network, inputting the data set two into the trained MLP network to obtain geometric features; inputting the data set two into the Encoder network to obtain feature vectors, fusing the feature vectors with the geometric features and then inputting them into the Decoder network, using the impedance spectrum as the output to train the Encoder-Decoder network based on LSTM, and constructing an impedance spectrum prediction model based on the trained MLP network and the trained Encoder-Decoder network; predicting the impedance spectrum; verifying the credibility of the impedance spectrum based on the Lin-KK method, and also providing a device for predicting the EIS of lithium batteries for power energy storage; solving the problems of insufficient physical interpretation, unclear feature construction, and low reliability existing in impedance spectrum prediction.

Description

Technical Field

[0001] The present invention relates to the technical field of lithium battery impedance spectrum prediction, and particularly to a method and device for predicting the EIS of lithium batteries for power energy storage. Background Art

[0002] Electrochemical Impedance Spectroscopy (EIS) data refers to a series of values of battery impedance obtained from battery impedance spectrum tests, which are used to study the electrochemical reactions inside the battery. It obtains the impedance spectrum of the battery by applying small-amplitude alternating current sinusoidal potential waves with different frequencies to the battery and measuring the ratio of the alternating potential to the current signal (i.e., the impedance of the system) as a function of the sinusoidal wave frequency. In actual research, the electrochemical performance of the battery, and even the evaluation of the battery life, etc., are often studied through the obtained battery impedance spectrum data. The impedance spectrum data of the battery is mainly obtained through professional impedance spectrum test instruments, but the impedance spectrum test instruments are relatively expensive, resulting in a high cost for obtaining impedance spectrum data. Therefore, how to obtain impedance spectrum data at low cost has become an urgent problem to be solved.

[0003] Meanwhile, with the rapid development of renewable energy, power energy storage technology plays an increasingly crucial role in regulating the fluctuations between power supply and demand, improving the stability of the power grid, and enhancing energy utilization efficiency. Lithium batteries, with their high energy density, long cycle life, and lightweight characteristics, have become one of the main battery technologies in power energy storage systems and are widely used in fields such as power system peak shaving, frequency modulation, off-grid systems, and electric vehicle charging stations. However, with the continuous expansion of the application scope of lithium batteries, the requirements for their performance monitoring, optimization, and life prediction are also increasing day by day. In particular, the internal resistance and impedance characteristics of the battery, as important parameters for evaluating the health status and life of lithium batteries, can obtain information such as internal electrochemical reactions, electron and ion transport in the battery by monitoring the impedance spectrum, which is of great significance for improving the intelligent level of the battery management system and extending the battery life. Therefore, how to obtain the impedance spectrum data of lithium batteries at low cost and high efficiency and accurately analyze their performance has become an important research direction in the research and application of power energy storage systems.

[0004] To address this, the currently common solution is as follows: establish a mapping relationship between the current data, voltage data, and temperature data of the battery and the impedance spectrum data, and construct a corresponding impedance spectrum prediction model. Just input the measured current data, voltage data, and temperature data into the impedance spectrum prediction model to predict the corresponding impedance spectrum data. Although this method can predict the impedance spectrum using charge-discharge data, it often has problems such as insufficient physical interpretation, unclear feature construction, and poor generalization ability. In addition, the current data-driven impedance spectrum prediction method does not introduce any mechanism to verify the reliability of the impedance spectrum prediction. If the target impedance spectrum in the training set of the data-driven model barely passes the KK test, then the model may introduce new uncertainties when generalizing to the test set, resulting in fluctuations in the reliability of the predicted impedance spectrum. Summary of the Invention

[0005] The technical problem to be solved by the present invention is how to solve the problems of insufficient physical interpretation, unclear feature construction, and low reliability in the current prediction of the impedance spectrum of lithium batteries.

[0006] The present invention solves the above technical problems through the following technical solutions: a method for predicting the EIS of a power energy storage lithium battery, the method comprising:

[0007] Divide the preprocessed battery sample data into a test set, a first data set, and a second data set, and calculate the geometric parameters of the battery in different states according to the impedance spectrum;

[0008] Use the first data set as the input and the geometric parameters as the output to train the MLP network to obtain the trained MLP network, and input the second data set into the trained MLP network to obtain geometric features;

[0009] Input the second data set into the Encoder network to obtain a feature vector. After fusing the feature vector with the geometric features, input it into the Decoder network, use the impedance spectrum as the output, train the Encoder-Decoder network based on LSTM to obtain the trained Encoder-Decoder network, and construct an impedance spectrum prediction model based on the trained MLP network and the trained Encoder-Decoder network;

[0010] Input the test set into the impedance spectrum prediction model to predict the impedance spectrum;

[0011] Verify the credibility of the impedance spectrum based on the Lin-KK method to obtain a credible impedance spectrum.

[0012] Beneficial effects: The present invention constructs an impedance spectrum prediction model using a two-stage network structure. In the first stage, by training an MLP network, a mapping relationship between battery sample data and geometric features is established to ensure that the geometric features output by the trained MLP network match the geometric morphology of the actual impedance spectrum. The geometric features closely related to the impedance spectrum morphology in the battery sample data are extracted through the trained MLP network, linking the battery charge and discharge behavior with the impedance spectrum morphology, making the prediction results have clear physical meanings. In the second stage, through deep mapping, the predicted physical geometric features and the battery sample data are jointly input into an Encoder-Decoder network. The sequential feature vectors of the battery sample data are extracted through the Encoder network, and after fusing the sequential feature vectors with the geometric features, they are input into the Decoder network to train the Encoder-Decoder network based on LSTM. According to the trained MLP network and the trained Encoder-Decoder network, an impedance spectrum prediction model is constructed. This model not only inherits the adaptability of the data-driven method to operating condition data but also establishes an interpretable association with the electrochemical mechanism through an explicit physical feature layer. Through the organic fusion of geometric features and sequential features, the prediction results have clear physical interpretability and high prediction accuracy, and can achieve the goal of directly mapping the battery sample data collected during the charge and discharge process to the full-frequency domain impedance spectrum of the battery. Finally, the credibility of the predicted impedance spectrum is verified based on an outlier removal algorithm verified by linear Kronig-Kramers, improving the reliability of impedance spectrum prediction.

[0013] Preferably, the battery sample data includes current, voltage, and temperature, and the preprocessing process includes:

[0014] Filter the collected battery sample data to obtain noise-removed battery sample data;

[0015] Normalize the noise-removed battery sample data to obtain normalized battery sample data;

[0016] Calibrate the normalized battery sample data to obtain preprocessed battery sample data.

[0017] Beneficial effects: The present invention filters the noise of the original signals of current, voltage, and temperature to remove interference signals. By normalizing the data, the influence of different measurement ranges on model training can be eliminated. By calibrating the data, the consistency and accuracy of the data at each sampling point can be ensured.

[0018] Preferably, the process of calculating the geometric parameters of the battery in different states according to the impedance spectrum includes:

[0019] Filter and fit the impedance spectrum to obtain a continuous EIS curve;

[0020] Based on the first derivative and the second derivative of the EIS curve, the boundary from high frequency to medium frequency and the boundary from medium frequency to low frequency are obtained respectively;

[0021] Select the abscissa of the intersection point of the high-frequency region and the real axis as a parameter , fit the corresponding semicircle equation in the medium-frequency region, and obtain the radius value of the circle as a parameter , fit the straight-line equation in the low-frequency region, and the slope of the straight-line equation is a parameter .

[0022] Beneficial effects: By analyzing the changing trends of the first and second derivatives of the EIS curve, the present invention determines the dividing positions of the low-frequency, medium-frequency, and high-frequency regions. When the first derivative drops sharply, it indicates that the EIS transitions from the high-frequency straight line to the medium-frequency semicircle. In the medium-frequency semicircle region, the change of the first derivative is relatively slow. When the second derivative shows a mutation, it corresponds to the first derivative changing from negative to positive, indicating the transition from the medium-frequency region to the low-frequency region. After dividing the frequency range, fit the medium-frequency semicircle to obtain the parameter , and at the same time extract the ohmic impedance corresponding to the low-frequency straight line and its slope angle , geometric parameters , , are the key parameters describing the geometric morphology of the EIS.

[0023] Preferably, the process of training the MLP network includes:

[0024] Input the data one into the input layer of the MLP network and flatten it to obtain a one-dimensional input vector;

[0025] Input the input vector into the first hidden layer of the MLP network to map it to a high-dimensional feature space;

[0026] Input the output of the first hidden layer into the second hidden layer for further mapping, and keep the output high-dimensional;

[0027] Input the output of the second hidden layer into the output layer to map and obtain geometric parameters;

[0028] Calculate the mean square error between the predicted value and the calculated value of the geometric parameter. When the mean square error is the smallest, the trained MLP network is obtained.

[0029] Beneficial effects: By training the MLP network, the present invention can ensure the physical meaning and stability of the extracted geometric features. During the training process, using the impedance spectrum data obtained by actual measurement, the network parameters are optimized through loss functions such as the mean square error (MSE) to ensure the output of g , r , The features match the geometric shape of the actual impedance spectrum. After training, the parameters of the geometric feature extraction module are fixed to directly provide stable and physically interpretable auxiliary information in subsequent stages. The present invention realizes the direct mapping from conventional BMS sensor data (voltage, current, temperature) to EIS geometric features ( g , r , ). Different from traditional methods that rely on dedicated impedance spectrometers to extract these features, the present invention directly predicts equivalent g , r , parameters from the operating condition data through a deep learning network, which not only retains the physical correlation between EIS features and battery aging mechanisms (such as g reflecting the ohmic impedance, r corresponding to the charge transfer process), but also gets rid of the dependence on high-frequency impedance detection equipment.

[0030] Preferably, the first hidden layer adopts the non-linear activation function , and the expression is:

[0031]

[0032] where is the input vector, and represents the value of the function.

[0033] Beneficial effect: Adopting the non-linear activation function Swish can provide a smoother gradient flow in the network, thereby accelerating the training process and avoiding the problem of gradient vanishing.

[0034] Preferably, the process of training the LSTM-based Encoder-Decoder network includes:

[0035] Inputting the second dataset into the trained MLP network to obtain geometric features;

[0036] Inputting the second dataset into the Encoder network to generate the feature vector ;

[0037] Concatenating the feature vector with the geometric features to obtain the comprehensive feature vector ;

[0038] Inputting the comprehensive feature vector into two Decoder networks respectively. One Decoder network outputs the real part corresponding to the impedance value at each frequency point, and the other Decoder network outputs the imaginary part corresponding to the impedance value at each frequency point. According to the real part and the imaginary part Obtain the predicted impedance spectrum;

[0039] Calculate the root mean square error between the predicted impedance spectrum and the actual impedance spectrum. When the root mean square error is minimized, the trained Encoder-Decoder network is obtained.

[0040] Preferably, the process of verifying the credibility of the impedance spectrum based on the Lin-KK method includes:

[0041] Obtain the experimental EIS data of the battery, set the maximum and minimum orders of the RC components in the Lin-KK model, calculate the fitting results of the experimental EIS data at different orders, and find the optimal order based on the fitting results;

[0042] Fit the impedance spectrum based on the Lin-KK model with the optimal order of the RC component to obtain the Lin-KK impedance of the impedance spectrum;

[0043] Calculate the residuals at each frequency point in the impedance spectrum based on the Lin-KK impedance, and eliminate the abnormal points with residuals greater than the set threshold to obtain a credible impedance spectrum.

[0044] Beneficial effects: When verifying the credibility of the predicted impedance spectrum in the present invention, first, the experimental EIS data from a high-precision electrochemical workstation is used to determine the optimal equivalent circuit RC component order of the current battery through the Lin-KK method, and this order represents the global topological characteristics of the overall impedance spectrum of the battery; then, the same order is maintained on the predicted impedance spectrum, and the remaining model parameters are finely fitted. By judging the residuals at each frequency point in the predicted impedance spectrum, only when the residuals are lower than the set threshold, the data is considered to highly match the true dynamic behavior of the electrochemical system at the local parameter level. Through the comprehensive verification of the two aspects of the physical internal consistency and the model topology and parameter consistency of the predicted impedance spectrum, not only the physical authenticity of the impedance spectrum is ensured, but also the stable consistency of the impedance spectrum in the global model structure is guaranteed, providing a solid data foundation for subsequent battery state assessment and accurate modeling.

[0045] Preferably, the process of calculating the fitting results of the experimental EIS data at different orders and finding the optimal order based on the fitting results is:

[0046] Set the initial value of the DC resistance , resistance , and time constant . At each order, simultaneously adjust the resistance and the time constant , and the value of the DC resistance is fixed or slightly adjusted within a small range. Iteratively calculate until the preset convergence condition is met to obtain the best fitting parameters at this order, and so on to obtain the best fitting parameters at all orders;

[0047] Based on the best - fit parameters at each order, calculate the real - part residuals of the experimental EIS data at each frequency point and the imaginary - part residuals : :

[0048]

[0049]

[0050] Calculate the average residual at each order M from the real - part residuals and the imaginary - part residuals :

[0051]

[0052] Take the order corresponding to the minimum average residual value as the optimal order :

[0053]

[0054] where, and are the real part and the imaginary part of the experimental EIS data at the frequency point respectively, and are the real part and the imaginary part of the impedance obtained by fitting with the Lin - KK model respectively, N is the number of frequency points.

[0055] Preferably, the Lin - KK impedance of the impedance spectrum is:

[0056]

[0057] where, and and are the DC resistance, the resistance of the th RC element, and the time constant of the k th RC element in the equivalent circuit after fitting when the order of the RC element is the optimal order k respectively;

[0058] The residuals of each frequency point in the impedance spectrum include the real - part residuals and the imaginary - part residuals :

[0059]

[0060]

[0061] Among them, and are respectively the real part and the imaginary part corresponding to the frequency points in the generated EIS data ; and are respectively the real part and the imaginary part of the impedance obtained by fitting the Lin-KK model when the order of the RC element is the optimal order;

[0062] When the real part residual or the imaginary part residual is greater than the set threshold, the data corresponding to this frequency point is determined as an abnormal point, and the abnormal point is removed from the impedance spectrum to obtain a credible impedance spectrum.

[0063] The present invention also provides an EIS prediction device for power energy storage lithium batteries, and the device includes:

[0064] A data processing module, configured to divide the preprocessed battery sample data into a test set, a data set one, and a data set two, and calculate the geometric parameters of the battery in different states according to the impedance spectrum;

[0065] A geometric feature extraction module, configured to use the data set one as the input and the geometric parameters as the output to train an MLP network to obtain a trained MLP network, and input the data set two into the trained MLP network to obtain geometric features;

[0066] A joint modeling module, configured to input the data set two into an Encoder network to obtain a feature vector, fuse the feature vector with the geometric features and then input them into a Decoder network, use the impedance spectrum as the output to train an Encoder-Decoder network based on LSTM to obtain a trained Encoder-Decoder network, and construct an impedance spectrum prediction model based on the trained MLP network and the trained Encoder-Decoder network;

[0067] A prediction module, configured to input the test set into the impedance spectrum prediction model to predict the impedance spectrum;

[0068] A verification module, configured to verify the credibility of the impedance spectrum based on the Lin-KK method to obtain a credible impedance spectrum.

[0069] The advantages provided by the present invention also include:

[0070] (1) The impedance spectrum prediction method of the present invention can online predict the full-frequency domain impedance spectrum according to the collected battery charge and discharge data, without the need for traditional impedance spectrum testing instruments and external perturbation signals, and can solve the problems of expensive testing equipment, long testing cycle, and insufficient physical interpretation of data existing in the prior art.

[0071] (2) Based on the fact that the Physics-Guided Neural Network (PGNN) of the present invention has a low computational latency, compared with complex models (such as graph neural networks), the PGNN network has a lower risk of overfitting for small-sample data, is more adaptable to the long-tailed distribution characteristics of battery aging data, and has stronger robustness. The two-stage network structure combines temporal dynamic features and geometric morphological features. The model not only inherits the adaptability of the data-driven method to operating condition data but also establishes an interpretable association with the electrochemical mechanism through an explicit physical feature layer (parameters parameters parameters ). The robustness and accuracy of the prediction are further improved through cross-validation and model fusion.

[0072] (3) By reusing the Lin-KK verification method, the present invention establishes a physical connection between the generated data and the experimental data. First, the RC order is initialized and set within a relatively wide range that basically covers the possible RC orders of the lithium battery equivalent circuit model. By fitting the experimental EIS data and calculating the real part residual and imaginary part residual at each order, the average residual is calculated based on the real part residual and imaginary part residual, and the order corresponding to the minimum average residual is determined as the optimal order. This optimal fitting order has the function of representing the basic information of the lithium battery equivalent circuit. When inspecting the quality of the predicted impedance spectrum, to ensure the consistency of the model under different operating conditions and facilitate comparison, the present invention fixes the RC element order of the model to the optimal order obtained by fitting, and then fits the remaining three parameters through the Lin-KK model, making the evaluation benchmark of the impedance spectrum consistent with that of the experimental EIS data, and successfully connecting the experimental EIS data with the predicted impedance spectrum data. BRIEF DESCRIPTION OF THE DRAWINGS

[0073] Figure 1 is a schematic diagram of geometric features in the method for predicting the impedance spectrum of a power storage lithium battery provided by an embodiment of the present invention;

[0074] Figure 2 is a schematic diagram of the principle of the method for predicting the impedance spectrum of a power storage lithium battery provided by an embodiment of the present invention;

[0075] Figure 3 is a flowchart of the method for predicting the impedance spectrum of a power storage lithium battery provided by an embodiment of the present invention;

[0076] Figure 4 is a sub-flowchart of the method for predicting the impedance spectrum of a power storage lithium battery provided by an embodiment of the present invention;

[0077] Figure 5 is a schematic diagram of the device for predicting the impedance spectrum of a power storage lithium battery provided by an embodiment of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0078] To make the objectives, technical solutions, and advantages of the present invention clearer and more understandable, the following describes the technical solutions of the present invention clearly and completely in conjunction with specific embodiments and with reference to the accompanying drawings. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without making creative efforts fall within the scope of protection of the present invention.

[0079] Embodiment 1

[0080] Refer to Figure 2 and Figure 3 , this embodiment provides a method for predicting the EIS of lithium batteries for power energy storage, including the following steps:

[0081] S101. During the charging and discharging process of the lithium-ion battery, collect multiple groups of battery sample data, preprocess the battery sample data to obtain the preprocessed battery sample data, and divide the preprocessed battery sample data into a test set, a data set one, and a data set two. For example, divide the preprocessed battery sample data into 10 parts, 1 part as the test set, and the remaining 9 parts as the data set for training and validating the model. The battery sample data includes current data, voltage data, and temperature data. The temperature data refers to the battery surface temperature signal. The data sampling rate is generally set to about 10 Hz, thereby forming a time series data matrix of size (3, T), where "3" represents three channels of voltage, current, and temperature, and T is the number of time steps within the sampling duration. During the process of obtaining each group of battery sample data, synchronously test the impedance value corresponding to the sample battery at a certain frequency through an impedance spectroscopy test instrument. The impedance spectroscopy data includes impedance values at different frequencies, and calculate the geometric parameters of the battery in different states according to the impedance spectroscopy.

[0082] The process of preprocessing the collected battery sample data includes:

[0083] Perform filtering on the battery sample data to remove interference signals and obtain the battery sample data with noise removed;

[0084] Perform normalization on the battery sample data with noise removed to obtain the normalized battery sample data to eliminate the influence of different measurement ranges on model training;

[0085] Perform calibration on the normalized battery sample data to obtain the preprocessed battery sample data to ensure the consistency and accuracy of the data at each sampling point. Through the preprocessing of the battery sample data, the high quality and stability of the input data of the subsequent network model can be guaranteed.

[0086] The process of calculating the geometric parameters of the battery in different states according to the impedance spectroscopy includes:

[0087] Input impedance spectrum (real part resistance values arranged from high to low in frequency and imaginary part resistance values ), filter and fit using a Savitzky-Golay filter to obtain a continuous EIS curve;

[0088] Analyze the changing trends of the first derivative and second derivative of the EIS curve to obtain the division positions of the low-frequency, mid-frequency, and high-frequency regions; among them, the first derivative reflects the change in curve curvature, the second derivative is used to capture inflection points, the first derivative rapidly drops from a large positive value to near zero (corresponding to the transition from a vertical line → semi-circle), and the physical meaning represented is that the ohmic impedance dominates → the charge transfer impedance dominates. The second derivative shows a local maximum value (corresponding to the end of the semi-circular arc and the start of the Warburg impedance), and the physical meaning represented is that the charge transfer impedance dominates → the diffusion process dominates. Based on the first derivative, the boundary from high frequency to mid frequency is obtained, and based on the second derivative, the boundary from mid frequency to low frequency is obtained.

[0089] Select the abscissa value of the intersection point of the high-frequency region of the EIS curve and the real axis as a parameter , use the least squares method to fit the semi-circle equation corresponding to the mid-frequency region to obtain the radius value of the circle as a parameter , use the real part and imaginary part data in the low-frequency region, and use the least squares method to fit a straight-line equation in the low-frequency region. The slope value of the straight-line equation is a parameter .

[0090] See Figure 1 , the geometric parameters , , These three parameters respectively correspond to the key geometric features of the impedance spectrum. Among them, the parameter is the voltage value corresponding to the intersection point of the impedance spectrum and the real axis in the Nyquist plot, indicating the ohmic resistance of the battery. The parameter is the fitting radius of the semi-circle in the mid-frequency region of the Nyquist plot, reflecting the interfacial reaction and related electrochemical processes. The parameter is the slope angle of the straight line in the low-frequency region of the Nyquist plot, reflecting the diffusion process and transport limitation effect.

[0091] S102. Using dataset one as the input and geometric parameters as the output, train the MLP network. When the loss function is minimized, obtain the trained MLP network. Input dataset two into the trained MLP network to obtain geometric features.

[0092] The data set used for training and verification in S101 can be divided into 9 parts, and 2 parts are selected as data set 1 for MLP network training. Data set 1 is used as input and geometric parameters are used as output. The MLP network is trained and the mean square error (MSE) is used as the loss function. When the loss function is minimized, the trained MLP network is obtained. The trained MLP network can extract geometric features that are highly consistent with the actual impedance spectrum data morphology from the input current, voltage, and temperature data. After the training is completed, the MLP network parameters are frozen for geometric feature generation in the second training stage.

[0093] The process of training an MLP network includes:

[0094] Input the battery sample data (two-dimensional data of size 3×T) into the input layer of the preset MLP network and flatten it to obtain a one-dimensional input vector;

[0095] The input vector is input into the first hidden layer of the preset MLP network and mapped to a high-dimensional feature space, such as a 256-dimensional feature space. The first hidden layer uses a nonlinear activation function (such as or ) Enhance the ability to express features and extract the electrochemical features and impedance spectrum geometry implicit in the charge and discharge data. Nonlinear activation function The mathematical expression is:

[0096]

[0097] in, is the input vector, express The value of the function.

[0098] Non-linear activation function It can provide a smoother gradient flow in the network, thereby speeding up the training process and avoiding the gradient vanishing problem.

[0099] The output of the first hidden layer is input into the second hidden layer for further mapping, so as to keep the output high-dimensional, and the dimension of the present invention is 256;

[0100] The output of the second hidden layer is input into the output layer, and the geometric parameters (parameters ,parameter ,parameter ), the equivalent parameters are predicted directly from the operating data through a deep learning network ,parameter ,parameter , which preserves the physical correlation between EIS characteristics and battery aging mechanisms (such as parameters Reflects ohmic impedance, parameters corresponding charge transfer process), and also gets rid of the dependence on high-frequency impedance detection equipment.

[0101] S103. Input the second dataset into the Encoder network to obtain a feature vector. After fusing the feature vector with geometric features, input it into the Decoder network, with the impedance spectrum as the output, and train the Encoder-Decoder network based on LSTM. When the loss function is minimized, obtain the trained Encoder-Decoder network. Build an impedance spectrum prediction model based on the trained MLP network and the trained Encoder-Decoder network.

[0102] Use the remaining 7 samples in the dataset for training and validation in S101 as the second dataset to train the Encoder-Decoder network. Input the battery sample data of the second dataset into the trained MLP network to obtain geometric features. Input the battery sample data into the Encoder network to generate a feature vector , the feature vector generated by the Encoder network represents the dynamic features of current, voltage, and temperature time series data, and can describe the feature vector of the dynamic changes in the battery charging and discharging process. The feature vector is concatenated with geometric features (parameters , parameters , parameters ) to obtain a comprehensive feature vector , the comprehensive feature vector is respectively input into two Decoder networks. One Decoder network outputs the real part corresponding to the impedance value at each frequency point , and the other Decoder network outputs the imaginary part corresponding to the impedance value at each frequency point . According to the real part and the imaginary part , the impedance spectrum EIS can be obtained.

[0103] The Encoder-Decoder network adopts an encoder-decoder structure based on an improved long short-term memory network for realizing the prediction of the full-frequency domain impedance spectrum. In order to improve the prediction performance of the model, when training the Encoder-Decoder network based on LSTM in the present invention, it is divided into two stages. In the first stage, jointly train the Encoder network and the two Decoder networks to obtain the preliminarily trained Encoder-Decoder network, and enhance the ability of the Decoder network to extract features from battery sample data. In the second stage, freeze the parameters of the Encoder network, and then independently train the two Decoder networks respectively. Based on the feature vector extracted by the encoder , improve the mapping ability of the decoder for the real and imaginary parts to obtain the trained Encoder-Decoder network. During the training process, the root mean square error (RMSE) is used as the loss function. When the loss function is minimized, the trained Encoder-Decoder network is obtained.

[0104] When training the MLP network and the LSTM-based Encoder-Decoder network, they can be trained using cross-validation and model fusion strategies, which can effectively alleviate the additional prediction errors that may be caused by the sample distribution deviation of a single model and improve the prediction accuracy and robustness of the model. For example, when training the MLP network, the first dataset is divided into a training set and a validation set according to a set ratio. The training set is used to train the model, and the model parameters are modified during the training process. The validation set is used to validate the trained model and adjust the hyperparameters of the model. In the first dataset, a group is sequentially selected as the validation set, and the rest are used as the training set to obtain multiple groups of datasets. Each group of datasets includes a group of validation sets and multiple groups of training sets. Training and validating the MLP network with each group of datasets can obtain multiple independent trained MLP networks. Similarly, when training the Encoder-Decoder network, the second dataset is divided into a training set and a validation set according to a set ratio. In the second dataset, a group is sequentially selected as the validation set, and the rest are used as the training set to obtain multiple groups of datasets. Each group of datasets includes a group of validation sets and multiple groups of training sets. Training and validating the Encoder-Decoder network with each group of datasets can obtain multiple independent trained Encoder-Decoder networks.

[0105] S104. Input the test set into the impedance spectrum prediction model to predict the impedance spectrum. When inputting the test set into the impedance spectrum prediction model, the current, voltage, and temperature data are input into multiple trained MLP networks to obtain multiple groups of output results. The multiple groups of output results are weighted and fused to obtain geometric features. The test set is input into the Encoder network of multiple independent trained Encoder-Decoder networks to obtain feature vectors. The feature vectors and geometric features are concatenated and input into the Decoder network to obtain multiple groups of output results. The multiple groups of output results are weighted and fused to obtain the predicted impedance spectrum, and the root mean square error is used to evaluate the prediction performance of the model. The root mean square error The calculation formula is:

[0106]

[0107] where, and respectively represent the real and imaginary parts of the actual EIS at the i th frequency point, and They are the real part and the imaginary part of the prediction result respectively.

[0108] An attention mechanism (Temporal Attention Mechanism) is introduced into the Encoder-Decoder network. This mechanism dynamically assigns different attention weights to each time step during the charge and discharge process of the lithium battery, enabling the model to focus on the most critical time periods during the charge and discharge process, so as to better pay attention to the battery behavior in the critical time periods and more accurately predict the impedance spectrum.

[0109] The attention weight of each time step The calculation formula is:

[0110]

[0111] Among them, is the attention score of each time step t , the larger it is, the more attention is paid to this time step. It is obtained by weighted calculation of the hidden state of the LSTM and the input feature :

[0112]

[0113] Among them, is usually the weight matrix for linear transformation of the hidden state or other hidden layer outputs, used to map the hidden vector to the same or additively compatible dimension as the input , b is the corresponding bias term, which is consistent with the usage of other biases in the neural network and is used to provide a learnable translation amount for the linear transformation, that is, and b are the weight matrix and bias term used for linear mapping of the hidden state or input vector respectively.

[0114] The present invention constructs an impedance spectrum prediction model using a two-stage network structure. In the first stage, by training an MLP network, a mapping relationship between battery sample data and geometric features is established to ensure that the geometric features output by the trained MLP network match the geometric shape of the actual impedance spectrum. The geometric features closely related to the impedance spectrum shape in the battery sample data are extracted through the trained MLP network, linking the battery charge and discharge behavior closely with the impedance spectrum shape, making the prediction results have clear physical meanings. In the second stage, through deep mapping, the predicted physical geometric features and the battery sample data are jointly input into an Encoder-Decoder network. The time series feature vector of the battery sample data is extracted through the Encoder network, and after fusing the time series feature vector with the geometric features, it is input into the Decoder network. The Encoder-Decoder network based on LSTM is trained, and an impedance spectrum prediction model is constructed according to the trained MLP network and the trained Encoder-Decoder network. This model not only inherits the adaptability of the data-driven method to operating condition data but also establishes an interpretable association with the electrochemical mechanism through an explicit physical feature layer. Through the organic fusion of geometric features and time series features, the prediction results have clear physical interpretability and high prediction accuracy, and can achieve the goal of directly mapping the battery sample data collected during the charge and discharge process to the full-frequency domain impedance spectrum of the battery.

[0115] The impedance spectrum prediction method of the present invention can online predict the full-frequency domain impedance spectrum based on the collected battery charge and discharge data, without the need for traditional impedance spectrum testing instruments and external perturbation signals, and can solve the problems existing in the prior art such as expensive testing equipment, long testing cycle, and insufficient physical interpretation of data. Compared with the traditional EIS prediction method, it does not need to interrupt the battery operation for impedance scanning and realizes real-time online monitoring; compared with the pure data-driven model, the present invention provides an explanation of the battery state change mode through the physical meanings of parameters 、 、 (such as a sudden increase in the value of parameter indicating the loss of active substances), which helps to improve the prediction accuracy; compared with the hybrid model, the present invention first uses the geometric features of the impedance spectrum as an intermediate variable rather than the final output, enabling the network to simultaneously learn physical laws and data distributions.

[0116] The PGNN network structure (with more than 80% fewer parameters than Transformer) has a lower computational delay. Compared with complex models (such as graph neural networks), the PGNN network has a lower overfitting risk for small sample data, is more adaptable to the long-tail distribution characteristics of battery aging data, and has stronger robustness. The two-stage network structure combines time series dynamic features and geometric shape features. The model not only inherits the adaptability of the data-driven method to operating condition data but also through an explicit physical feature layer (parameter 、parameter , Parameter ) An interpretable association with the electrochemical mechanism is established, and the robustness and accuracy of the prediction are further improved through cross-validation and model fusion. The present invention is applicable to the prediction of impedance spectra of various types of lithium-ion batteries. Especially in the fields of electric vehicles, energy storage systems, etc., it can realize real-time online monitoring of battery impedance spectra and contribute to the construction of a precise battery management system.

[0117] S105. Verify the credibility of the impedance spectrum based on the Lin-KK method to obtain a credible impedance spectrum. Refer to Figure 4 , and the specific process of credibility verification includes:

[0118] S1051. Obtain the experimental EIS data of the battery. The experimental EIS data is obtained by testing the battery with a high-precision electrochemical workstation under test standards. The experimental EIS data includes frequency data , real part data , and imaginary part data . Among them, the frequency data includes a series of positive frequency points, the real part data is the real part corresponding to each positive frequency point, and the imaginary part data is the imaginary part corresponding to each positive frequency point.

[0119] Set the maximum order and minimum order of the RC elements in the Lin-KK model, calculate the fitting results of the experimental EIS data at different orders, and find the optimal order based on the fitting results.

[0120] Among them, the impedance of the Lin-KK model is:

[0121]

[0122] Among them, is the DC resistance in the equivalent circuit of the lithium battery, is the resistance of the k th RC element in the equivalent circuit of the lithium battery, is the time constant of the k th RC element in the equivalent circuit of the lithium battery, M is the order of the RC element in the equivalent circuit of the lithium battery.

[0123] The present invention sets the maximum order of the RC element in the Lin-KK model to 50, and the minimum order is 1. In other embodiments, the model order can be simplified. For example, for some data, if the contribution of high-order RC elements is small, the calculation efficiency can be improved by simplifying the model order (selecting more appropriate maximum and minimum orders). The least squares method is used to fit the experimental EIS data at different orders. The fitting process includes:

[0124] Set the initial value of the DC resistance , resistance , time constant . For example, the average value of the experimental EIS data in the high-frequency region can be used as the initial value of the DC resistance . Select a small positive value as the initial value of the resistance . Select a fixed value between 1 ms and 10 ms as the initial value of the time constant . It can also be judged by analyzing the data characteristics.

[0125] Order M ranges from the minimum order to the maximum order . In this embodiment, the maximum order is 50, and the minimum order is 1. For each selected order M , while adjusting the resistance and the time constant , the value of the DC resistance remains fixed or is slightly adjusted within a small range. Iterative calculation is performed until the preset convergence condition is met, and the best fitting parameters at this order are obtained. The preset convergence condition can be that the change in the model parameters to be fitted (DC resistance , resistance , time constant ) is less than the set threshold, or the number of iterations reaches the upper limit value. By analogy, the best fitting parameters at all orders are obtained. The best fitting parameters include the DC resistance , resistance , time constant .

[0126] Based on the best fitting parameters at each order, calculate the real part residual and the imaginary part residual of the experimental EIS data at each frequency point :

[0127]

[0128]

[0129] Calculate the average residual at each order M according to the real part residual and the imaginary part residual :

[0130]

[0131] Take the order corresponding to the minimum average residual value as the optimal order :

[0132]

[0133] Among them, is the experimental EIS data, , are respectively the real part and the imaginary part corresponding to the frequency point in the experimental EIS data, , are respectively the real part and the imaginary part of the impedance obtained by fitting through the Lin-KK model, N is the number of frequency points.

[0134] S1052. Fit the impedance spectrum obtained in step S106 with the Lin-KK model based on the optimal order of the RC element to obtain the Lin-KK impedance of the impedance spectrum; the impedance spectrum includes frequency data , real part data , and imaginary part data . Among them, the frequency data includes a series of positive frequency points, the real part data is the real part corresponding to each frequency point, and the imaginary part data is the imaginary part corresponding to each frequency point.

[0135] Select the order of the RC element as the optimal order obtained in the above step S1051 , and use the least squares method, such as the Levenberg-Marquardt algorithm, to fit the impedance spectrum obtained in step S104. The fitting method is similar to the fitting process in the above step S1051, that is, first set the initial values of the DC resistance , resistance , and time constant . For example, the average value of the impedance spectrum in the high-frequency region can be used as the initial value of the DC resistance , a small positive value can be selected as the initial value of the resistance , and a fixed value between 1 ms and 10 ms can be selected as the initial value of the time constant . It can also be judged by analyzing the characteristics of the generated impedance spectrum. During the fitting process, adjust the resistance and the time constant simultaneously, and the value of the DC resistance remains fixed or is slightly adjusted within a small range. Iteratively calculate until the preset convergence condition is met to obtain the optimal order The best fitting parameters under , , .

[0136] Lin-KK impedance of impedance spectrum for:

[0137]

[0138] in, , , The order of RC components is the optimal order. When the DC resistance and the k The resistance of the first RC element, k The time constant of an RC element.

[0139] The present invention basically covers the possible RC orders of the lithium battery equivalent circuit model within a wide range (1-50), and determines the optimal equivalent circuit RC element order of the current battery through the Lin-KK method. This order represents the global topological characteristics of the overall impedance spectrum of the battery, that is, the optimal fitting order has the function of representing the basic information of the lithium battery equivalent circuit. When testing the quality of the predicted impedance spectrum, in order to ensure the consistency of the model under different working conditions and facilitate comparison, the present invention fixes the RC element order of the model to the optimal order obtained by the above fitting, and then fits the remaining three parameters through the Lin-KK model, so that the evaluation benchmark of the predicted impedance spectrum is consistent with the evaluation benchmark of the experimental EIS data, and successfully links the experimental EIS data with the data-driven predicted impedance spectrum.

[0140] S1053. Calculate the residual of each frequency point in the impedance spectrum based on the Lin-KK impedance, remove abnormal points whose residual is greater than a set threshold, and obtain a reliable impedance spectrum.

[0141] The residual at each frequency point in the impedance spectrum includes the real part residual , imaginary residual :

[0142]

[0143]

[0144] in, To generate EIS data, , The frequency points in the generated EIS data are The corresponding real and imaginary parts are , They are the real part and the imaginary part of the impedance obtained by fitting the Lin-KK model when the order of the RC element is the optimal order, respectively.

[0145] When the real part residual or the imaginary part residual is greater than the set threshold In this embodiment, the set threshold is 0.01 or 0.001. The data corresponding to this frequency point is determined as an abnormal point, and the abnormal point is removed from the impedance spectrum to obtain a credible impedance spectrum. That is, only when the real part residual and the imaginary part residual are satisfied, the data at this frequency point is considered qualified. In other embodiments, the set threshold can be adaptively adjusted according to different experimental conditions and data quality to further improve the inspection accuracy.

[0146] After screening the predicted impedance spectrum, the screened EIS data is output as a credible impedance spectrum. The screened EIS data removes unqualified frequency points, conforms to the standard of the Lin-KK model, and can be used for further analysis or modeling.

[0147] The present invention verifies the credibility of the impedance spectrum based on the Lin-KK method. First, using the experimental EIS data from a high-precision electrochemical workstation, the optimal equivalent circuit RC element order of the current battery is determined by the Lin-KK method, and this order represents the global topological characteristics of the overall impedance spectrum of the battery. Then, on the predicted impedance spectrum, the same order is maintained, and the remaining model parameters are finely fitted. By judging the residuals of each frequency point in the predicted impedance spectrum, only when the residuals are lower than the set threshold, it is considered that the data highly coincides with the true dynamic behavior of the electrochemical system at the local parameter level. Through the comprehensive verification of the physical internal consistency and the model topology and parameter consistency of the predicted impedance spectrum, not only the physical authenticity of the impedance spectrum is ensured, but also the stable consistency of the impedance spectrum in the global model structure is guaranteed, providing a solid data basis for subsequent battery state assessment and accurate modeling.

[0148] Embodiment 2

[0149] Refer to Figure 5 , this embodiment provides a lithium battery EIS prediction device for power energy storage. The device includes:

[0150] A data processing module, configured to divide the preprocessed battery sample data into a test set, a data set one, and a data set two, and calculate the geometric parameters of the battery in different states according to the impedance spectrum; the battery sample data includes current, voltage, and temperature. The process of preprocessing the collected battery sample data includes:

[0151] Filter the collected battery sample data to obtain the battery sample data with noise removed;

[0152] Normalize the battery sample data with noise removed to obtain the normalized battery sample data;

[0153] Calibrate the normalized battery sample data to obtain the preprocessed battery sample data.

[0154] The process of calculating the geometric parameters of the battery in different states according to the impedance spectrum includes:

[0155] Filter and fit the impedance spectrum to obtain a continuous EIS curve;

[0156] Based on the first derivative and the second derivative of the EIS curve, obtain the boundary from high frequency to medium frequency and the boundary from medium frequency to low frequency respectively;

[0157] Select the abscissa of the intersection point of the high-frequency region and the real axis as a parameter , fit the corresponding semicircle equation in the medium-frequency region, and obtain the radius value of the circle as a parameter , fit the straight-line equation in the low-frequency region, and the slope of the straight-line equation is a parameter .

[0158] The geometric feature extraction module is used to train the MLP network with Dataset 1 as the input and geometric parameters as the output. When the loss function is minimized, the trained MLP network is obtained. Input Dataset 2 into the trained MLP network to obtain geometric features. The process of training the MLP network includes:

[0159] Input Data 1 into the input layer of the MLP network and flatten it to obtain a one-dimensional input vector;

[0160] Input the input vector into the first hidden layer of the MLP network to map it to a high-dimensional feature space;

[0161] Input the output of the first hidden layer into the second hidden layer for further mapping, and keep the output high-dimensional;

[0162] Input the output of the second hidden layer into the output layer to map and obtain geometric parameters;

[0163] Calculate the mean square error between the predicted value and the calculated value of the geometric parameters. When the mean square error is minimized, the trained MLP network is obtained.

[0164] The first hidden layer uses the non-linear activation function Swish, and the expression is:

[0165] .

[0166] The joint modeling module is used to input Dataset 2 into the Encoder network to obtain a feature vector. After the feature vector is fused with the geometric features, it is input into the Decoder network, with the impedance spectrum as the output, to train the LSTM-based Encoder-Decoder network. When the loss function is minimized, the trained Encoder-Decoder network is obtained. The impedance spectrum prediction model is constructed based on the trained MLP network and the trained Encoder-Decoder network. The process of training the LSTM-based Encoder-Decoder network includes:

[0167] Input Dataset 2 into the trained MLP network to obtain geometric features;

[0168] Input Dataset 2 into the Encoder network to generate a feature vector ;

[0169] Concatenate the feature vector with the geometric features to obtain a comprehensive feature vector ;

[0170] Input the comprehensive feature vector into two Decoder networks respectively. One Decoder network outputs the real part corresponding to the impedance value at each frequency point , and the other Decoder network outputs the imaginary part corresponding to the impedance value at each frequency point . According to the real part and the imaginary part , obtain the predicted impedance spectrum;

[0171] Calculate the root mean square error between the predicted impedance spectrum and the actual impedance spectrum. When the root mean square error is minimized, the trained Encoder-Decoder network is obtained.

[0172] The prediction module is used to input the test set into the impedance spectrum prediction model to predict the impedance spectrum;

[0173] The verification module is used to verify the credibility of the impedance spectrum based on the Lin-KK method to obtain a credible impedance spectrum.

[0174] The process of verifying the credibility of the impedance spectrum based on the Lin-KK method includes:

[0175] Obtain the experimental EIS data of the battery, set the maximum and minimum orders of the RC elements in the Lin-KK model, calculate the fitting results of the experimental EIS data at different orders, and find the optimal order based on the fitting results;

[0176] Fit the impedance spectrum based on the Lin-KK model with the optimal order of the RC elements to obtain the Lin-KK impedance of the impedance spectrum;

[0177] Calculate the residuals of each frequency point in the impedance spectrum based on the Lin-KK impedance, and eliminate the abnormal points with residuals greater than the set threshold to obtain a reliable impedance spectrum.

[0178] Among them, the process of calculating the fitting results of the experimental EIS data at different orders and finding the optimal order based on the fitting results is as follows:

[0179] Set the initial value of the DC resistance , resistance , and time constant . At each order, simultaneously adjust the resistance and time constant , and keep the value of the DC resistance fixed or slightly adjusted within a small range. Iteratively calculate until the preset convergence condition is met to obtain the best fitting parameters at this order, and so on to obtain the best fitting parameters at all orders;

[0180] Based on the best fitting parameters at each order, calculate the real part residual and imaginary part residual of the experimental EIS data at each frequency point :

[0181]

[0182]

[0183] Calculate the average residual M at each order according to the real part residual and imaginary part residual:

[0184]

[0185] Take the order corresponding to the minimum average residual value as the optimal order :

[0186]

[0187] Among them, , are the real part and imaginary part corresponding to the frequency point in the experimental EIS data , , are the real part and imaginary part of the impedance obtained by fitting through the Lin-KK model, N is the number of frequency points.

[0188] The Lin-KK impedance of the impedance spectrum is:

[0189]

[0190] Among them, 、 、 are respectively the DC resistance, the resistance of the -th RC element, and the time constant of the k -th RC element in the equivalent circuit after fitting when the order of the RC element is the optimal order k ;

[0191] The residuals at each frequency point in the impedance spectrum include the real part residual and the imaginary part residual :

[0192]

[0193]

[0194] Among them, 、 are respectively the real part and the imaginary part corresponding to the frequency point in the generated EIS data , 、 are respectively the real part and the imaginary part of the impedance obtained by fitting with the Lin-KK model when the order of the RC element is the optimal order;

[0195] When the real part residual or the imaginary part residual is greater than the set threshold, the data corresponding to this frequency point is determined as an abnormal point, and the abnormal point is removed from the impedance spectrum to obtain a reliable impedance spectrum.

[0196] Example 3

[0197] This example takes a dataset containing charge-discharge and EIS data of 11 brand-new cylindrical lithium iron phosphate batteries as the research object.

[0198] Each battery is subjected to a constant current (CC) discharge test under different states (SOC from 5% to 100%, step size 5%). The test conditions include a discharge current of 0.12 A, a cut-off voltage of 2 V, and an ambient temperature of 20°C. The capacity of each battery is between 633 mAh and 683 mAh. For each battery, an electrochemical impedance spectroscopy (EIS) test is performed under different SOC conditions. The test frequency range is from 0.01 Hz to 1000 Hz, and 28 sets of real part and imaginary part data are obtained for each test. For each EIS test, the first 100 current, voltage, and temperature data points are intercepted from the discharge process corresponding to the SOC as input data, and a total of 220 data samples are finally formed. Among them, battery B07 is randomly selected as the test sample, and the remaining data is used for model training and verification.

[0199] To achieve online prediction, in this embodiment, the original current, voltage, and temperature time-series data collected during the charge and discharge process are first preprocessed. The preprocessing steps include:

[0200] (1) Using a digital filtering algorithm to suppress noise in the original data and remove high-frequency interference;

[0201] (2) Normalizing the current, voltage, and temperature data according to their respective measurement ranges;

[0202] (3) Correcting the time-series data during the acquisition process to ensure the consistency of the data in the time domain.

[0203] The preprocessed data forms a matrix of size (3, 100), where "3" represents the three channels of voltage, current, and temperature, providing high-quality input for subsequent feature extraction and prediction.

[0204] Next, the preprocessed data is input into the two-stage physics-guided neural network (PGNN) model of the present invention for full-frequency domain EIS prediction. In the first stage, a pre-trained geometric feature extraction module is used. This module is based on a multi-layer perceptron (MLP) network structure and uses the Swish activation function to enhance the feature expression ability. The flattened (3, 100) data is converted into the key parameters , parameter , parameter . The specific implementation steps are as follows:

[0205] (1) Flatten the (3, 100) data into a one-dimensional vector;

[0206] (2) After passing through two hidden layers (both set to 256 neurons and using the Swish activation function), three scalars are generated through the output layer: parameter , parameter , parameter , where parameter reflects the ohmic internal resistance corresponding to the intersection point of the EIS with the real axis in the Nyquist plot, parameter reflects the radius of the semicircle fitting in the intermediate frequency region, representing the interfacial electrochemical reaction, and parameter reflects the slope of the straight line in the low-frequency region, characterizing the diffusion and transport limitation effects.

[0207] In this stage, the geometric parameters at each SOC are pre-calculated using the actually measured EIS data, and the mean square error (MSE) is used as the loss function to train the MLP until the extracted geometric features highly match the actual EIS morphology. After training, the parameters of this module are frozen and used as the physical constraint for the second-stage prediction.

[0208] In the second stage, an encoder-decoder structure based on an improved long short-term memory network (LSTM) is adopted, and a temporal attention mechanism is introduced. This mechanism calculates attention weights for each time step, enabling the model to focus on the features of key time periods during the charge and discharge process, thereby improving the prediction accuracy:

[0209] (1) Input the preprocessed voltage and current time series data into a single-layer LSTM encoder to extract its temporal dynamic features and generate feature vectors ;

[0210] (2) Concatenate the feature vectors output by the encoder with the geometric features obtained in the first stage ( , , ) to form a comprehensive feature vector ;

[0211] (3) Use two independent LSTM decoders to process the comprehensive features respectively. One decoder is responsible for predicting the real part of the EIS, and the other decoder is responsible for predicting the imaginary part;

[0212] (4) Each of the two decoders maps the output to the impedance values at each frequency point through a fully connected layer, and combines the real part and imaginary part results to obtain the complete predicted EIS.

[0213] During the training process, the AdamW optimizer (with an initial learning rate set to 0.001 and combined with a dynamic learning rate decay strategy) is adopted, and at the same time, a gradient clipping strategy is introduced to jointly train the encoder and decoder parameters, with the goal of minimizing the RMSE between the predicted EIS and the actual measured EIS at each frequency point.

[0214] To further improve the robustness and generalization ability of the model, this embodiment adopts a cross-validation and model fusion strategy. The specific implementation method is as follows:

[0215] (1) Among the 11 battery samples, fix battery B07 as the test set, and sequentially select 1 from the remaining 10 battery samples as the validation set, and the rest as the training set to form multiple data partitioning schemes;

[0216] (2) Train the model for each data partitioning scheme respectively to obtain multiple independent trained models;

[0217] (3)Integrate the EIS prediction results of each model in the test set under the same SOC condition by means of weighted fusion to form the final prediction output. This strategy effectively alleviates the prediction deviation of a single model that may be caused by the difference in sample distribution, and significantly improves the overall prediction accuracy and stability.

[0218] For the evaluation of the prediction performance, the root mean square error (RMSE) is used as the main evaluation index in this embodiment. Through multiple experimental statistics, under all SOC conditions, the overall EIS prediction RMSE of the test set is lower than 1 mΩ, and the prediction errors in each frequency band are maintained at a low level, which proves the high precision and stability of the method of the present invention.

[0219] The experimental results show that under the conditions of SOC = 30% and 55%, the predicted EIS curves are highly consistent with the actual measurement data at most frequency points; among them, the prediction accuracy of the real part is relatively high, and there are slight deviations in the imaginary part in the low-frequency region, but the overall error is controlled within the allowable range. Further, by statistically analyzing the RMSE distribution under different SOCs, the results show that the median RMSE under each SOC condition is lower than 1 mΩ, and the maximum error does not exceed 0.1 Ω, which fully verifies the effectiveness of this method under multiple working conditions.

[0220] This embodiment verifies through experiments on 11 battery samples that the full-frequency EIS online prediction method based on a physics-guided neural network can effectively utilize the voltage and current data collected during the charge and discharge process to achieve high-precision prediction of the impedance spectrum of lithium-ion batteries. This method has the advantages of online real-time, low cost, simple structure, good physical interpretability and robustness, and is suitable for practical applications of battery state monitoring and health management in electric vehicles, energy storage systems and battery management systems.

[0221] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements on some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. EIS prediction method for lithium batteries for power storage, characterized by: Methods include: The preprocessed battery sample data is divided into a test set, data set 1, and data set 2. The geometric parameters of the battery under different states are calculated based on the impedance spectrum, including: The impedance spectrum is filtered and fitted to obtain a continuous EIS curve; Based on the first-order derivative and second-order derivative of the EIS curve, the boundary from high frequency to medium frequency and the boundary from medium frequency to low frequency are obtained respectively; Select the horizontal coordinate of the intersection of the high frequency area and the real axis as the parameter , fitting the semicircle equation corresponding to the mid-frequency region, the radius of the circle is the parameter , fit the straight line equation in the low frequency region, the slope of the straight line equation is the parameter ; Taking data set 1 as input and geometric parameters as output, the MLP network is trained to obtain the trained MLP network. Data set 2 is input into the trained MLP network to obtain geometric features. The second dataset is input into the Encoder network to obtain the feature vector, which is then fused with the geometric features and input into the Decoder network. The impedance spectrum is used as the output to train the LSTM-based Encoder-Decoder network to obtain the trained Encoder-Decoder network. The impedance spectrum prediction model is constructed based on the trained MLP network and the trained Encoder-Decoder network. The test set is input into the impedance spectrum prediction model to predict the impedance spectrum; The credibility of the impedance spectrum is verified based on the Lin-KK method, and a reliable impedance spectrum is obtained.

2. The EIS prediction method for lithium battery for power energy storage according to claim 1, characterized in that: Battery sample data includes current, voltage, and temperature. The preprocessing process includes: Performing filtering processing on the collected battery sample data to obtain battery sample data with noise removed; Normalizing the battery sample data after noise removal to obtain normalized battery sample data; The normalized battery sample data is corrected to obtain preprocessed battery sample data.

3. The EIS prediction method for lithium battery for power energy storage according to claim 1, characterized in that: The process of training an MLP network includes: Input the data into the input layer of the MLP network and flatten it to obtain a one-dimensional input vector; Input the input vector to the first hidden layer of the MLP network and map it to a high-dimensional feature space; The output of the first hidden layer is input into the second hidden layer for further mapping to keep the output high dimensionality; The output of the second hidden layer is input into the output layer to map the geometric parameters; The mean square error between the predicted value and the calculated value of the geometric parameters is calculated. When the mean square error is minimized, the trained MLP network is obtained.

4. The EIS prediction method for lithium battery for power energy storage according to claim 3, characterized in that: The first hidden layer uses a nonlinear activation function , the expression is: in, is the input vector, express The value of the function.

5. The EIS prediction method for lithium battery for power energy storage according to claim 1, characterized in that: The process of training an LSTM-based Encoder-Decoder network includes: Input the trained MLP network into the second dataset to obtain geometric features; Input data set 2 into the Encoder network to generate feature vector ; The feature vector Combined with geometric features to obtain comprehensive feature vector ; The comprehensive feature vector Input two decoder networks respectively, and one decoder network outputs the real part of the impedance value corresponding to each frequency point , another Decoder network outputs the imaginary part of the impedance value at each frequency point , according to the real part and the imaginary part Get the predicted impedance spectrum; The root mean square error between the predicted impedance spectrum and the actual impedance spectrum is calculated. When the root mean square error is the smallest, the trained Encoder-Decoder network is obtained.

6. The EIS prediction method for lithium battery for power energy storage according to claim 1, characterized in that: The process of verifying the credibility of impedance spectrum based on Lin-KK method includes: Obtain the experimental EIS data of the battery, set the maximum and minimum orders of the RC elements in the Lin-KK model, calculate the fitting results of the experimental EIS data under different orders, and find the optimal order based on the fitting results; The impedance spectrum is fitted based on the Lin-KK model with the optimal order of RC elements to obtain the Lin-KK impedance of the impedance spectrum; The residual of each frequency point in the impedance spectrum is calculated based on Lin-KK impedance, and the abnormal points with residuals greater than the set threshold are eliminated to obtain a reliable impedance spectrum.

7. The EIS prediction method for lithium battery for power energy storage according to claim 6, characterized in that: The fitting results of experimental EIS data under different orders are calculated. The process of finding the optimal order based on the fitting results is as follows: Setting DC resistance ,resistance , time constant The initial value of the resistor is adjusted at each order. and time constant , DC resistance The value of is fixed or fine-tuned within a small range, and the iterative calculation is performed until the preset convergence condition is met to obtain the best fitting parameters under this order, and the best fitting parameters under all orders are obtained by analogy; Based on the best fitting parameters at each order, the experimental EIS data are calculated at each frequency point The real residual and the imaginary residual : Calculate each order based on the real and imaginary residuals M The average residual under : The order corresponding to the minimum average residual value is taken as the optimal order : in, , The experimental EIS data are Mid-frequency point The corresponding real and imaginary parts are , are the real and imaginary parts of the impedance obtained by fitting the Lin-KK model, N is the number of frequency points.

8. The EIS prediction method for lithium battery for power energy storage according to claim 6, characterized in that: Lin-KK impedance of impedance spectrum for: in, , , The order of RC components is the optimal order. When the DC resistance and the k The resistance of the first RC element, k The time constant of an RC element; The residual at each frequency point in the impedance spectrum includes the real part residual , imaginary residual : in, , Generate EIS data for Mid-frequency point The corresponding real and imaginary parts are , They are respectively the real part and imaginary part of the impedance obtained by fitting the Lin-KK model when the order of the RC element is the optimal order; When the real residual or the imaginary residual When it is greater than the set threshold, the data corresponding to the frequency point is determined as an abnormal point, and the abnormal point is removed from the impedance spectrum to obtain a reliable impedance spectrum.

9. EIS prediction device for lithium batteries for power storage, characterized by: The device includes: The data processing module is used to divide the pre-processed battery sample data into a test set, a data set 1, and a data set 2, and calculate the geometric parameters of the battery under different states according to the impedance spectrum, including: The impedance spectrum is filtered and fitted to obtain a continuous EIS curve; Based on the first-order derivative and second-order derivative of the EIS curve, the boundary from high frequency to medium frequency and the boundary from medium frequency to low frequency are obtained respectively; Select the horizontal coordinate of the intersection of the high frequency area and the real axis as the parameter , fitting the semicircle equation corresponding to the mid-frequency region, the radius of the circle is the parameter , fit the straight line equation in the low frequency region, the slope of the straight line equation is the parameter ; The geometric feature extraction module is used to train the MLP network with the data set 1 as input and the geometric parameters as output to obtain the trained MLP network, and input the data set 2 into the trained MLP network to obtain the geometric features; The joint modeling module is used to input the second data set into the Encoder network to obtain a feature vector, which is then fused with the geometric features and input into the Decoder network. The impedance spectrum is used as the output to train the LSTM-based Encoder-Decoder network to obtain a trained Encoder-Decoder network. The impedance spectrum prediction model is constructed based on the trained MLP network and the trained Encoder-Decoder network. A prediction module is used to input the test set into the impedance spectrum prediction model to predict the impedance spectrum; The verification module is used to verify the credibility of the impedance spectrum based on the Lin-KK method to obtain a credible impedance spectrum.

10. The EIS prediction device for lithium battery for electric energy storage according to claim 9, characterized in that: The process of training an MLP network includes: Input the data into the input layer of the MLP network and flatten it to obtain a one-dimensional input vector; Input the input vector to the first hidden layer of the MLP network and map it to a high-dimensional feature space; The output of the first hidden layer is input into the second hidden layer for further mapping to keep the output high dimensionality; The output of the second hidden layer is input into the output layer to map the geometric parameters; The mean square error between the predicted value and the calculated value of the geometric parameters is calculated. When the mean square error is minimized, the trained MLP network is obtained.

Images