Lithium battery rapid data extrapolation method based on machine learning dual-channel agent model

Abstract

The application discloses a lithium battery rapid data extrapolation method based on a machine learning double-channel agent model, relates to a lithium battery health state evaluation method and aims to solve the problems of low calculation efficiency and poor real-time performance of an existing electrochemical model in voltage curve prediction. The method comprises the following steps: refining high-sensitivity parameters, combining a charge-discharge rate to generate a parameter set, extracting a parameter combination, inputting the parameter combination into an electrochemical model, collecting a voltage curve, extracting time characteristics and shape characteristics after interpolation to construct a mapping relationship between parameters and characteristics, combining reverse interpolation to restore a complete voltage curve, using a multilayer perception model to train a mapping relationship of overpotential under a fixed exchange current density and a reaction current density, replacing an overpotential calculation module in a traditional electrochemical model, using a double-channel agent model and a particle swarm optimization algorithm to perform parameter identification on partial discharge data, inputting an identification result into the multilayer perception model to replace a partial electrochemical model, and completing charge-discharge curve output extrapolation. The method is used for battery voltage curve prediction.

Description

Technical Field

[0001] This invention belongs to the field of power battery modeling and health status assessment technology, specifically involving a method for rapid extrapolation of lithium-ion battery voltage curves based on a dual-channel surrogate model and parameter optimization. Background Technology

[0002] With the rapid development of new energy vehicles and energy storage systems, lithium-ion batteries, as their core energy unit, have attracted widespread attention for performance prediction technology. Traditional methods rely on electrochemical models based on physical mechanisms to simulate and invert voltage curves. Although these methods have high physical consistency, the calculation process is complex, and parameter identification requires a large amount of offline data and iterative optimization, making it difficult to meet the real-time and engineering deployment requirements of online applications.

[0003] In recent years, machine learning techniques have been introduced into battery modeling and health estimation tasks, enabling more efficient data-driven predictions by constructing a mapping relationship between battery operating characteristics and state variables. However, purely data-driven methods often lack physical interpretability and generalization ability, especially with decreased prediction accuracy outside the sampling range or under non-standard operating conditions, limiting their large-scale application. Therefore, there is an urgent need for a modeling method that can balance physical rationality, modeling efficiency, and generalization ability, enabling rapid analysis of partial battery discharge data and reliable extrapolation of complete voltage curves, providing more practical technical support for SOH assessment and lifetime prediction. Summary of the Invention

[0004] This invention aims to address the problems of low computational efficiency and poor real-time performance in the parameter identification and voltage curve prediction processes of existing electrochemical models, and provides a method for rapid extrapolation of lithium-ion battery voltage curves that combines dual-channel surrogate modeling and particle swarm optimization algorithm.

[0005] The fast data extrapolation method for lithium batteries based on a machine learning dual-channel surrogate model of the present invention includes the following steps:

[0006] Step 1: Using parameter sensitivity analysis, highly sensitive electrochemical parameters are screened from the electrochemical model. Combined with different charge and discharge rate conditions, a parameter combination dataset is constructed. The parameter combination is extracted using the Latin hypercube sampling method and substituted into the electrochemical model to construct a voltage curve dataset.

[0007] Step 2: Extract the time features of the voltage curve and the shape features of the interpolated curve. Use the least squares gradient boosting regression tree model to train the mapping relationship between the electrochemical parameter combination and the time features. Use the convolutional neural network model to train the mapping relationship between the electrochemical parameter combination and the shape features. Restore the output voltage curve result of the electrochemical model under a fixed parameter combination by back interpolation.

[0008] Step 3: Use a multilayer perceptron model to train the mapping relationship between the overpotential under a fixed exchange current density and the reaction current density, replacing the overpotential calculation module in the traditional electrochemical model to improve computational efficiency.

[0009] Step 4: Parameter identification is performed on partial discharge data using a dual-channel surrogate model and particle swarm optimization algorithm. The identification results are input into a multilayer perceptron for partial substitution of the electrochemical model, completing the output extrapolation of the charge-discharge curve. This invention is used for rapid extrapolation of voltage curve results.

[0010] Preferably, in step 1, the electrochemical parameters are divided into three groups: positive electrode electrochemical parameters, electrolyte electrochemical parameters, and negative electrode electrochemical parameters. The four parameters with the highest sensitivity among the three groups are selected as high-sensitivity parameters, and the sensitivity calculation is expressed as follows:

[0011] ; The average value of the parameter scan range. The upper bound of the parameter scan, The lower bound for parameter scanning, This is the root mean square error.

[0012] Preferably, the output voltage results of the electrochemical model ensure that the time characteristics satisfy: the charging cutoff voltage of the output results is 4.3V and the discharging cutoff voltage is 3V for each parameter combination.

[0013] Preferably, the simulation conditions of the electrochemical model are: first, constant current discharge to 3V, rest for 120s, and then constant current charging to 4.3V.

[0014] Preferably, the charge / discharge rate is kept consistent under the same set of parameters, and the charge / discharge rate range is 0.5C~1.5C. The parameter range extracted by Latin hypercube is consistent with the parameter optimization range of particle swarm optimization algorithm.

[0015] Preferably, in step 2, all data in the voltage curve dataset are interpolated to a fixed number of points of 300 to ensure the consistency of the training dimensions of the convolutional neural network.

[0016] Preferably, in step 3, the exchange current density and reaction current density are sampled within the range that may occur during actual simulation operation, and the upper and lower limits of sampling are determined in combination with the distribution of electrochemical parameters.

[0017] Preferably, in step 4, the data used for parameter identification is as small as possible while ensuring that sufficient physical information is covered, and the data segment of the parameter charging and discharging switching area is not selected.

[0018] Preferably, in order to maximize the utilization of data segment features, the particle swarm optimization algorithm uses multiple error indices for fitness calculation. The calculation process includes the root mean square error of the entire data segment, the absolute interpolation of the data segment termination voltage, and the absolute interpolation of the data segment start voltage. The weighting coefficients are determined based on the actual numerical range of the three error indices.

[0019] Preferably, the fitness of the particle swarm optimization algorithm described in step 4 is expressed as:

[0020] In the formula, , , These represent the weighting coefficients of the three error indicators. This is the simulation termination time. This is the end time of the experiment. To simulate the starting voltage, This is the starting voltage for the experiment.

[0021] The beneficial effects of this invention are:

[0022] This invention, based on a dual-channel surrogate model and particle swarm optimization algorithm, achieves rapid identification of electrochemical parameters using only partial discharge data. It also incorporates a multilayer perceptron to replace some computational modules in the electrochemical model, enabling efficient extrapolation of the complete voltage curve. By constructing surrogate models with two independent channels—time and shape features—the modeling task of complex voltage curves is effectively decomposed, reducing model training difficulty and improving prediction accuracy and controllability. Compared to traditional electrochemical modeling methods based on full data identification, this invention significantly reduces the required input data and identification computational overhead, possessing stronger engineering practicality and online deployment capabilities. Furthermore, this invention combines physical consistency with modeling efficiency, maintaining high prediction accuracy under different rate or parameter extrapolation conditions, verifying its good generalization ability and application value. Attached Figure Description

[0023] Figure 1 This is a flowchart of the fast data extrapolation method for lithium batteries based on a machine learning dual-channel proxy model according to the present invention.

[0024] Figure 2 This is a simplified electrochemical model of the lithium-ion battery in Example 1;

[0025] Figure 3 The flowchart shows the training method of the dual-channel proxy model used in Example 1.

[0026] Figure 4 This is a diagram illustrating the extrapolation effect of rapid data extrapolation in Example 1 under different proportions of partial data.

[0027] Figure 5The graph shows the extrapolation effect of rapid data extrapolation under different magnification conditions in Example 1. Detailed Implementation

[0028] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. The present invention will be further described below with reference to the accompanying drawings and specific embodiments, but this is not intended to limit the present invention.

[0029] Example 1: The fast data extrapolation method for lithium batteries based on a machine learning dual-channel surrogate model in this example is performed according to the following steps:

[0030] Step 1: Using parameter sensitivity analysis, highly sensitive parameters are screened from the electrochemical model. Combined with different charge / discharge rate conditions, a parameter combination dataset is constructed. The Latin hypercube sampling method is then used to extract parameter combinations, which are then substituted into the electrochemical model to construct a voltage curve dataset. Details are as follows:

[0031] In step 1, to ensure that the positive electrode parameters, negative electrode parameters, and electrolyte parameters are all included in the database generation process, the highly sensitive parameters are selected as follows: all electrochemical parameters are divided into three groups: positive electrode-related electrochemical parameters, negative electrode-related electrochemical parameters, and electrolyte-related electrochemical parameters. From these three groups, the four parameters with the highest sensitivity within each group are selected, for a total of twelve parameters, as highly sensitive parameters. The sensitivity calculation formula is as follows:

[0032] In the formula: The average value of the parameter scan range. The upper bound of the parameter scan, The lower bound for parameter scanning, This represents the root mean square error. The scan range for each parameter is shown in Table 1, where... This represents the initial concentration of the electrolyte. This represents the maximum lithium-ion concentration in the positive electrode solid phase particles. The electrolyte diffusion coefficient is... This represents the volume fraction of the solid phase at the positive electrode. The thickness of the positive electrode coating. The thickness of the electrolyte region. The radius of the positive electrode particle is... The thickness is the negative electrode thickness. For the SEI film conductivity, The lithium concentration in the solid phase of lithium metal. For the SEI film thickness, is the coefficient for lithium insertion / extraction reaction at the positive electrode.

[0033] Table 1: Range of values ​​for highly sensitive parameters

[0034]

[0035] In step 1, to ensure that the model output results have consistent time characteristics and termination conditions, specific constraints were set for the voltage output of the electrochemical model: under each parameter combination, the output voltage curve must meet the following conditions: discharge cutoff voltage of 3.0V and charging cutoff voltage of 4.3V. The specific simulation conditions are set as follows: the battery initially discharges at a constant current until the voltage drops to 3.0V; then it enters a resting phase with a resting time of 120s; finally, it is charged at a constant current until the voltage rises to 4.3V. Through these standardized operations, the consistency of all simulation data across the time axis and voltage range is ensured, which is beneficial for subsequent comparative analysis and learning modeling of model behavior under different parameter combinations.

[0036] In step 1, to ensure the comparability and consistency of simulation data under different rate conditions, the charge / discharge rate is always kept consistent under the same set of parameters. That is, the charging and discharging processes corresponding to each set of parameters use the same rate setting. The rate range is set from 0.5C to 1.5C, covering typical low-to-medium rate application scenarios. Simultaneously, to ensure the consistency and comparability of the parameter space, the parameter range covered by the Latin hypercube sampling method is strictly aligned with the parameter optimization boundary in the particle swarm optimization algorithm used for fitting real data, and is consistent with the parameter scanning range during the determination of highly sensitive parameters. Details are shown in Table 1, ensuring the representativeness of the modeling dataset and the consistency of the parameter identification results.

[0037] In step 1, the value range of each parameter is divided into multiple equally probable sub-intervals, with the number of sub-intervals matching the number of samples in the database. A sampling point is randomly selected within each sub-interval to ensure that samples for each dimension are uniformly distributed across its entire range. By independently sampling each dimension and then randomly combining them, multiple unique multidimensional parameter sets are generated as input samples for the electrochemical model. This method effectively avoids the problem of sample aggregation or omission in high-dimensional space, improves the coverage efficiency of the parameter space, and provides a high-quality data foundation for subsequent efficient simulation and surrogate modeling of voltage curves.

[0038] 2500 representative parameter combinations were generated in the parameter space. The interval for each parameter was divided into 2500 equally probable sub-intervals, and a value was randomly selected from each sub-interval to ensure that the value of each parameter was uniformly distributed within its range. Then, the 2500 sampled values ​​of the 12 parameters were randomly combined to form 2500 sets of 12-dimensional parameter samples. Finally, these 2500 sets of samples were sequentially input into the electrochemical model to generate corresponding voltage curve data for training the surrogate model.

[0039] Step 2: Extract the time features and interpolated curve shape features of the voltage curve. Use a least-squares gradient boosting regression tree model to train the mapping relationship between the electrochemical parameter combination and the time features. Use a convolutional neural network model to train the mapping relationship between the electrochemical parameter combination and the shape features. Finally, reconstruct the output voltage curve of the electrochemical model with a fixed parameter combination using back-interpolation. Details are as follows:

[0040] In step 2, feature decomposition is performed on each original voltage curve to extract time and shape features. Time features refer to the total discharge time experienced by the battery under the current parameter combination, from the initial voltage of 4.3V to the final voltage of 3V, i.e., the end value of the time series corresponding to the complete voltage curve. This feature reflects the impact of parameters on the overall capacity release process. Shape feature extraction is achieved by linear interpolation of the original voltage curves, uniformly resampling the unequal-length original curves into equally spaced curves of a fixed length (i.e., 300 points), preserving their normalized relative shape information. This process ensures the alignment and comparability of voltage change processes under different parameter combinations and facilitates the input of shape information as a fixed-length vector into a convolutional neural network for modeling, ensuring the consistency of convolutional neural network training. The least squares gradient boosting regression tree model is used to train the mapping relationship between electrochemical parameter combinations and time features, and the convolutional neural network model is used to train the mapping relationship between electrochemical parameter combinations and shape features. The output voltage curve result of the electrochemical model under a fixed parameter combination is restored through back interpolation.

[0041] Step 3: Use a multilayer perceptron model to train the mapping relationship between the overpotential under a fixed exchange current density and the reaction current density, replacing the overpotential calculation module in the traditional electrochemical model to improve computational efficiency, as detailed below:

[0042] In step 3, based on the idea of ​​machine learning substitution, to improve the computational efficiency of the electrochemical model in voltage curve simulation and parameter identification, a multilayer perceptron model is used to replace the overpotential calculation module in the electrochemical model. Traditional electrochemical models require solving the Butler–Volmer (B–V) equation at each time step and spatial node to obtain the overpotential of the electrode reaction. This equation is a highly nonlinear implicit expression, typically relying on Newton's method or other numerical iterative algorithms for point-by-point solution. This results in high computational cost and slow speed, severely limiting the model's application in high-resolution simulations, large-scale parameter scans, or online predictions. This numerical solution process is replaced by a prediction process directly mapped by a machine learning model.

[0043] In step 3, the exchange current density and reaction current density are sampled within the range that may occur during actual simulation operation, and the upper and lower limits of sampling are determined in conjunction with the distribution of electrochemical parameters. The corresponding actual overpotential values ​​are calculated using traditional numerical calculation methods, thereby constructing a training dataset of exchange current density, reaction current density, and overpotential as the output. This input-output relationship is then fitted using several fully connected layers and nonlinear activation functions to obtain an efficient prediction model that can output the overpotential without iteration.

[0044] Step 4: Use a dual-channel surrogate model and particle swarm optimization algorithm to identify parameters for some discharge data. Input the identification results into a multilayer perceptron to perform partial substitution of the electrochemical model and complete the output extrapolation of the charge-discharge curve.

[0045] In step 4, the particle swarm optimization algorithm is used to efficiently invert key parameters of the electrochemical model under the constraint of partial discharge data. Traditional electrochemical models typically require complex numerical simulations and iterative processes to repeatedly adjust model parameters during parameter identification, resulting in high computational overhead and low fitting efficiency, making it difficult to meet the needs of online or rapid application scenarios. Therefore, an alternative structure based on dual-channel surrogate modeling is introduced to achieve rapid extrapolation of lithium battery data based on the dual-channel surrogate model. The specific process is as follows: Figure 3 As shown, the discharge duration and voltage curve shape of the electrochemical model are modeled separately, and trained using a least-squares gradient boosting regression tree and a convolutional neural network respectively, thus quickly generating predicted curves without requiring a complete simulation. Based on this, a particle swarm optimization (PSO) algorithm uses combinations of electrochemical parameters as particle codes to iteratively search in a multi-dimensional parameter space. After each particle update, the parameter set represented by the current particle is input into a dual-channel surrogate model to generate the corresponding complete voltage curve prediction result, which is then compared with some measured curves. The fitness function of the PSO algorithm is calculated as follows:

[0046] In the formula, , , These represent the weighting coefficients of the three error indicators. This is the simulation termination time. This is the end time of the experiment. To simulate the starting voltage, The starting voltage is set to the experimental voltage. This function is designed as a multi-objective weighted combination, including three metrics: the root mean square error of the entire curve, used to measure the overall fitting accuracy; the absolute error between the predicted starting voltage and the measured initial value, used to constrain the consistency of the model's starting point; and the difference between the predicted total discharge duration and the target duration, used to ensure the accuracy of the curve's time scale. These three errors are weighted and integrated through weighting coefficients to form a comprehensive fitness function, guiding the particle swarm to achieve a balance between physical consistency and numerical accuracy. After iterative convergence, the obtained optimal parameter combination is input into an electrochemical model containing a multilayer perceptron surrogate module to achieve extrapolation and reconstruction of the complete voltage curve.

[0047] In step 4, the proportion of data extracted needs to be determined. This means deciding what percentage of the entire discharge data should be extracted for parameter identification. The data used for parameter identification should be as short as possible while ensuring sufficient physical information is covered, avoiding data segments from the charge-discharge switching region. Specifically, if the extracted proportion is too small, such as less than 10%, the voltage change segment may be too gradual, making it difficult to demonstrate the modulating effect of parameters on dynamic behavior, leading to an under-constrained state in the particle swarm optimization process and unstable parameter identification. Conversely, if the extracted proportion is too large, such as above 70%, while it improves identification accuracy, it also increases data acquisition time and computational cost, weakening the advantage of rapid response. If only data segments from the charge-discharge switching region are used, excessively large data abrupt changes are also detrimental to the accurate extraction of data features. To achieve a balance between accuracy and efficiency, experimental comparative analysis is needed. Under multiple representative combinations of electrochemical parameters and rate conditions, the parameter inversion effects of different data ratios (e.g., 10%, 20%, 40%, 70%) are systematically evaluated. The three characteristics—RMSE, initiation voltage error, and discharge duration error—are integrated into a comprehensive index according to weighting coefficients, and this comprehensive index is used as the evaluation standard. The results are as follows: Figure 4 As shown, when selecting the first 20% of the discharge data, the predicted curve generated by the surrogate model exhibits high consistency in global shape, start point alignment, and termination duration, and its computational cost is significantly lower than that of full curve inversion. Therefore, this embodiment uses the first 20% of the entire discharge curve as partial data input, ensuring identification accuracy while also meeting real-time requirements, thus improving the feasibility and flexibility of the method in practical applications.

[0048] Step 4 also includes: To fully verify the generalization ability and practicality of the constructed dual-channel surrogate model under different rate conditions, a rate extrapolation evaluation scheme based on real experimental data was designed. Specifically, during the training phase, a dual-channel surrogate model was constructed using a large amount of parameter-voltage curve data generated by the electrochemical model at 1C rate. This model includes a regression model for predicting discharge duration and a neural network model for predicting the shape of the normalized curve. To evaluate the adaptability of the surrogate model under non-training rate conditions, two rates within the sampling range (0.75C and 1.25C) and two rates outside the sampling range (0.25C and 1.75C) were selected as validation conditions. For these four rate conditions, the voltage-time curves measured in real experiments were used as the evaluation benchmark, no longer relying on the simulation output of the electrochemical model, to ensure that the evaluation results have higher engineering reference value. During the testing process, partial discharge data at various discharge rates were first input into a particle swarm optimization algorithm. A set of optimal electrochemical parameters was then derived using a surrogate model. Subsequently, this parameter set was input into the trained surrogate model to predict the discharge duration and normalized voltage shape. Inverse interpolation was then used to recover the complete voltage curve, which was compared with the experimentally measured voltage at the corresponding discharge rates. By comparing the root mean square error, onset voltage error, and termination duration error between the predicted and measured curves at different discharge rates, the results are as follows: Figure 5 As shown, at the sampling rates (0.75C and 1.25C), the model's predicted curves and experimental curves are in high agreement, with good consistency in overall trend, start and end voltages, and duration, verifying the model's fitting ability. At the sampling rates (0.25C and 1.75C), although the error increases relatively, the model still maintains a physically reasonable discharge trend and low overall error, demonstrating a certain rate extrapolation capability. These results fully demonstrate that the surrogate model constructed in this embodiment not only has high accuracy at the training rates but also possesses good cross-rate generalization ability, adapting to the complex operating conditions of frequent rate changes in practical applications, providing reliable support for the rapid assessment and prediction of battery health status.

[0049] This embodiment decomposes the voltage curve into two sub-channels: time-related features and shape-related features. Mapping relationships between electrochemical parameters and various features are established separately, enabling efficient and accurate substitution and restoration of the electrochemical model output. A hybrid model is proposed, utilizing a multilayer perceptron to establish a mapping relationship between exchange current density, reaction current density, and overpotential, replacing the overpotential calculation module in the electrochemical model and further improving simulation efficiency. The dual-channel proxy model is combined with a particle swarm optimization algorithm to achieve rapid identification of highly sensitive electrochemical parameters. The identification results are input into the hybrid model to complete efficient and rapid extrapolation of the complete voltage curve, solving the problem of low computational efficiency and inability to meet online application requirements in existing electrochemical models during voltage curve prediction and parameter identification.

[0050] While the invention has been described herein with reference to specific embodiments, it should be understood that these embodiments are merely examples of the principles and applications of the invention. Therefore, it should be understood that many modifications can be made to the exemplary embodiments, and other arrangements can be designed without departing from the spirit and scope of the invention as defined by the appended claims. It should be understood that different dependent claims and features described herein can be combined in ways different from those described in the original claims. It is also understood that features described in conjunction with individual embodiments can be used in other described embodiments.

Claims

1. A fast data extrapolation method for lithium batteries based on a machine learning dual-channel surrogate model, characterized in that, The method includes the following steps: Step 1: Using parameter sensitivity analysis, highly sensitive electrochemical parameters are screened from the electrochemical model. Combined with different charge and discharge rate conditions, a parameter combination dataset is constructed. The parameter combination is extracted using the Latin hypercube sampling method and substituted into the electrochemical model to construct a voltage curve dataset. Step 2: Extract the time features of the voltage curve and the shape features of the interpolated curve. Use the least squares gradient boosting regression tree model to train the mapping relationship between the electrochemical parameter combination and the time features. Use the convolutional neural network model to train the mapping relationship between the electrochemical parameter combination and the shape features. Restore the output voltage curve result of the electrochemical model under a fixed parameter combination by back interpolation. Step 3: Use a multilayer perceptron model to train the mapping relationship between the overpotential under a fixed exchange current density and the reaction current density, replacing the overpotential calculation module in the traditional electrochemical model to improve computational efficiency. Step 4: Use a dual-channel surrogate model and particle swarm optimization algorithm to identify parameters for some discharge data. Input the identification results into a multilayer perceptron to perform partial substitution of the electrochemical model and complete the output extrapolation of the charge-discharge curve.

2. The method for fast data extrapolation of lithium batteries based on a dual-channel surrogate model of machine learning according to claim 1, characterized in that, In step 1, the electrochemical parameters are divided into three groups: positive electrode electrochemical parameters, electrolyte electrochemical parameters, and negative electrode electrochemical parameters. The four parameters with the highest sensitivity from the three groups are selected as high-sensitivity parameters. The sensitivity calculation is expressed as follows: In the formula: The average value of the parameter scan range. The upper bound of the parameter scan, The lower bound for parameter scanning, This is the root mean square error.

3. The method for fast data extrapolation of lithium batteries based on a dual-channel surrogate model of machine learning according to claim 1 or 2, characterized in that, In step 1, the output voltage result of the electrochemical model ensures that the time characteristics are met: the charging cutoff voltage of the output result is 4.3V and the discharging cutoff voltage is 3V under each parameter combination; the simulation conditions of the electrochemical model are: first constant current discharge to 3V, rest for 120s, and then constant current charging to 4.3V.

4. The method for fast data extrapolation of lithium batteries based on a dual-channel surrogate model of machine learning according to claim 1 or 2, characterized in that, In step 1, the charge and discharge rates under the same set of parameters are kept consistent, with the charge and discharge rate range being 0.5C to 1.5C. The parameter range extracted by Latin hypercube is consistent with the parameter optimization range of the particle swarm algorithm.

5. The method for fast data extrapolation of lithium batteries based on a dual-channel surrogate model of machine learning according to claim 1 or 2, characterized in that, In step 2, all data in the voltage curve dataset are interpolated to a fixed number of points of 300 to ensure the consistency of the training dimensions of the convolutional neural network.

6. The method for fast data extrapolation of lithium batteries based on a machine learning dual-channel surrogate model according to claim 1 or 2, characterized in that, In step 3, the exchange current density and reaction current density are sampled within the range that may occur during the actual simulation operation, and the upper and lower limits of sampling are determined in combination with the distribution of electrochemical parameters.

7. The method for fast data extrapolation of lithium batteries based on a dual-channel surrogate model of machine learning according to claim 1 or 2, characterized in that, The particle swarm optimization algorithm uses multiple error indices for fitness calculation. The calculation process includes the root mean square error of the entire data segment, the absolute interpolation of the data segment termination voltage, and the absolute interpolation of the data segment start voltage. The weighting coefficients are determined based on the actual numerical range of the three error indices.

8. The method for fast data extrapolation of lithium batteries based on a machine learning dual-channel surrogate model according to claim 7, characterized in that, The fitness of the particle swarm optimization algorithm is represented as: The fitness of the particle swarm optimization algorithm described in step 4 is expressed as: In the formula, , , These represent the weighting coefficients of the three error indicators. This is the simulation termination time. This is the end time of the experiment. To simulate the starting voltage, This is the starting voltage for the experiment.

Images