A model training method, a lithium battery evaluation model and device

Abstract

This invention belongs to the field of lithium battery state assessment technology. Addressing the technical bottlenecks of traditional physical models in terms of multi-scale coupling and computational efficiency, this invention proposes a model training method, a lithium battery assessment model, and a device. The lithium battery assessment model includes an encoder-neural network solver-decoder architecture. This invention constructs a hidden state vector H to uniformly represent the internal dynamics of the battery and employs neural network models such as Neural ODE and the Koopman operator to solve the hidden state evolution, overcoming the shortcomings of low computational efficiency and cumbersome modeling processes of traditional numerical calculation models.

Description

Technical Field

[0001] This invention belongs to the field of lithium battery state assessment technology, specifically relating to a model training method, a lithium battery assessment model, and a device. Background Technology

[0002] Globally, the goal of reducing carbon emissions has brought lithium-ion batteries (LIBs) into the spotlight. With their high energy density and extended lifespan, LIBs have become a key component of modern energy storage solutions. In 2024, the global LIB market reached a remarkable 1202.6 GWh, a year-on-year increase of 25.6%. However, the exponential expansion of applications has highlighted LIB degradation mechanisms and reliability issues, creating a critical bottleneck for the advancement of energy storage systems. This environment makes detailed and rapid LIB model simulations a crucial part of identifying the optimal design and operating modes for energy storage systems.

[0003] The evolution of lithium-ion battery modeling theory can be traced back to the field of electrochemical kinetics. The pseudo-two-dimensional (P2D) model developed by Doyle and Newman established a mathematical framework by coupling solid-liquid phase diffusion, potential distribution, and Butler-Volmer kinetics. To improve computational efficiency, the single-particle model (SPM) simplified the electrolyte concentration distribution at low rates, while its enhanced version (SPMe) incorporated electrolyte kinetics to broaden its applicability. Recent research has led to the development of a comprehensive set of modeling tools: the integration of P2D models with thermal and mechanical constitutive equations. Therefore, these multiphysics models are indispensable for thermal management and structural integrity studies.

[0004] The current field of lithium battery modeling faces many technical bottlenecks. For example, the construction of common pseudo-2D electrochemical models (P2D) requires the coupled solution of solid-liquid phase diffusion equations, potential distribution equations, and Butler-Volmer reaction kinetic equations, which results in single-cell charge-discharge simulations taking more than several hours, failing to meet real-time monitoring requirements, and the computational complexity increases exponentially when the battery pack is expanded.

[0005] Patent CN120428118A discloses a method and system for dynamic SOE prediction and lifetime assessment of lithium iron phosphate battery packs integrating a P2D model. Based on historical operating data and real-time state parameters, the invention uses a P2D model coupled with electrochemical-thermodynamic multiphysics fields to generate dynamic SOE prediction results for the battery pack. It then analyzes the prediction results in conjunction with long-term performance degradation mechanisms to assess the remaining lifetime of the battery pack and outputs the assessment results. This improves SOE prediction accuracy and comprehensively reflects performance change patterns using a lifetime assessment model. Through real-time monitoring and optimization, it significantly improves operating efficiency and extends service life. While simplified models, such as equivalent circuit models, have high computational efficiency, they cannot reflect the internal state of the battery and achieve cross-scale coupled calculations.

[0006] Patent CN120405448A discloses a method, apparatus, and medium for predicting the state of charge (SOC) of a lithium battery. The method includes: acquiring operational data of the lithium battery to be tested; inputting the acquired operational data into a pre-constructed SOC prediction model to obtain the predicted SOC of the lithium battery to be tested, as predicted by the SOC prediction model; the SOC prediction model predicts the SOC based on an extended Kalman filter framework, and updates the error covariance matrix of each iteration using the Levenberg-Marquardt method within the extended Kalman filter framework, and performs each iteration by numerically adjusting the iteration coefficients based on the influence of the Levenberg-Marquardt method during iteration. This method can only perform the single function of SOC prediction, lacks interpretability, and cannot estimate the corresponding state of the lithium battery electrodes. Summary of the Invention

[0007] This invention provides a model training method, a lithium battery evaluation model, and a device, aiming to solve the technical bottlenecks of traditional physical models in terms of multi-scale coupling and computational efficiency, and to achieve accurate simulation of the entire scale from electrode micro-dynamics to battery pack macro-behavior, providing a technical foundation for battery health status assessment and safety management.

[0008] The first aspect of this invention provides a model training method, comprising: Acquire historical data of lithium battery state variables within a time window; A lithium battery evaluation model based on neural networks was built using the TensorFlow framework. The target lithium battery evaluation model is obtained by training a neural network-based lithium battery evaluation model using historical data of lithium battery state variables. The target lithium battery evaluation model is used to evaluate the internal state of the lithium battery under test based on the state variable data of the lithium battery under test.

[0009] Furthermore, the neural network-based lithium battery evaluation model includes an encoder-neural network solver-decoder architecture; The encoder is used to encode the input historical data of lithium battery state variables; The neural network solver, located between the encoder and the decoder, is used to merge the encoded historical data of the lithium battery state variables and the hidden state vector of the previous time step, and to calculate the hidden state vector of the current time step. The decoder is used to decode the hidden state vector calculated at the current time step and output lithium battery state variable evaluation data.

[0010] Furthermore, the formula for calculating the hidden state vector at the current time step is: ; Among them, H t Let H be the state vector of the lithium battery at time t. t-1 Let x be the state vector of the lithium battery at time t-1. t To load the condition vector.

[0011] Furthermore, the state variables of a lithium battery include the initial state and boundary conditions; The initial state includes battery voltage, battery state of charge, positive / negative electrode potential, positive / negative electrode interface overpotential, average lithium concentration of positive / negative electrode particles, lithium concentration on the surface of positive / negative electrode particles, and electrolyte potential. Boundary conditions include lithium battery current and temperature.

[0012] Furthermore, the neural network solver can be any one of the following: GRU-based solver, Koopman operator evolution solver, or Neural ODE solver.

[0013] Furthermore, both the encoder and decoder employ a multilayer perceptron structure.

[0014] Furthermore, the multilayer perceptron structure consists of two fully connected layers; The first fully connected layer has a ReLU activation function; the second fully connected layer has a linear function.

[0015] Furthermore, the dimension of the hidden state vector is 8, and the number of hidden state vectors is 128.

[0016] A second aspect of this invention provides a lithium battery evaluation method, comprising the following steps: Acquire the state variable data of the lithium battery under test; The state variable data of the lithium battery to be tested is input into the target lithium battery evaluation model, and the lithium battery state variable evaluation data is output to evaluate the internal state of the lithium battery to be tested; wherein, the target lithium battery evaluation model is obtained based on any of the model training methods described above; A third aspect of the present invention provides a lithium battery evaluation apparatus, comprising: The data acquisition module is used to acquire the state variable data of the lithium battery under test; The evaluation module is used to input the state variable data of the lithium battery to be tested into the target lithium battery evaluation model and output the estimated state variable data of the lithium battery to be tested, thereby evaluating the internal state of the lithium battery to be tested; wherein, the target lithium battery evaluation model is obtained based on any of the model training methods described above.

[0017] A fourth aspect of the present invention provides an electronic device including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor executes the computer program to implement the steps of the model training method or the lithium battery evaluation method described above.

[0018] A fifth aspect of the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the model training method or the lithium battery evaluation method described above.

[0019] Compared with the prior art, the advantages of this invention are as follows: This invention constructs a hidden state vector H to uniformly represent the internal dynamic process of the battery, and uses neural network models such as Neural ODE and Koopman operator to solve the evolution of the hidden state, overcoming the shortcomings of traditional numerical calculation models such as low computational efficiency and cumbersome modeling process. Attached Figure Description

[0020] Figure 1 This is a framework diagram of the neural network-based lithium battery evaluation model in this invention; Figure 2 The structure diagrams of the five candidate models are shown below; (a) is a simple neural network; (b) is a two-layer neural network; (c) is a GRU-based solver model; (d) is a Koopman operator model; and (e) is a neural ODE model. Figure 3 This is a flowchart of the computation process of the neural network solver (solver) in this invention; Figure 4 A flowchart of the lithium battery evaluation method provided by the present invention; Figure 5The training results and prediction performance of the neural network-based lithium battery evaluation model are shown in the figure; (a) is the training loss curve; (b) is the mean absolute error of the prediction; (c) is the voltage and positive potential; (d) is the negative potential / overpotential, positive overpotential and electrolyte potential; (e) is the state of charge (SoC), negative / positive average / surface SoC; (f) is the magnitude of the model error when the input Gaussian noise error is used. Figure 6 The error of the neural network-based lithium battery evaluation model under long sequence prediction and three improved methods is shown in the figure. Among them, (a) is the average MAE of the neural ODE model with different units; (b) is the parameter-average MAE plot; (c) is a schematic diagram of the physical variable correction method; (d) is the average MAE of the neural ODE model after physical variable correction; and (e) is the average MAE of the neural ODE model learned from samples with different sequence lengths. Detailed Implementation

[0021] The following will be described in conjunction with embodiments of the present invention. Figures 1 to 6 The technical solutions in the embodiments of the present invention are clearly and completely described herein. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0022] Example 1 This invention provides a model training method, specifically including: Acquire historical data of lithium battery state variables within a time window; Specifically, for example / experimental scenarios: historical data of lithium battery state variables such as voltage, current, temperature, and possible internal states such as electrode potential are obtained through sensors; for numerical model / surrogate calculation model scenarios: historical data of lithium battery state variables are obtained through high-precision numerical simulation model parameter calculation. A lithium battery evaluation model based on neural networks was built using the TensorFlow framework. A target lithium battery evaluation model is obtained by training a neural network-based lithium battery evaluation model using historical lithium battery state variable data; the target lithium battery evaluation model is used to evaluate the internal state of the lithium battery under test based on the state variable data of the lithium battery under test.

[0023] In this embodiment of the invention, the lithium battery evaluation model based on neural networks (GRU) includes an encoder-neural network solver-decoder architecture; The encoder is used to encode the input historical data of lithium battery state variables; The neural network solver, located between the encoder and the decoder, is used to merge the historical data of the encoded lithium battery state variables and the hidden state vector of the previous time step, and to calculate the hidden state vector of the current time step. The decoder is used to decode the hidden state vector calculated at the current time step and output lithium battery state variable evaluation data.

[0024] In this embodiment of the invention, the formula for calculating the hidden state vector at the current time step is: ; Among them, H t Let H be the state vector of the lithium battery at time t. t-1 Let x be the state vector of the lithium battery at time t-1. t To load the condition vector.

[0025] In this embodiment of the invention, the state variables of the lithium battery include the initial state and boundary conditions; The initial state includes battery voltage, battery state of charge, positive / negative electrode potential, positive / negative electrode interface overpotential, average lithium concentration of positive / negative electrode particles, surface lithium concentration of positive / negative electrode particles, and electrolyte potential; the initial state is directly inherited from the output of the previous decoder step. Boundary conditions include lithium battery current and temperature.

[0026] In practical implementation, in order for the model to effectively identify and update the internal state of the battery while maintaining the underlying conservation laws, the encoder input can incorporate multi-frame fusion data of lithium battery state variables within a time window.

[0027] In this embodiment of the invention, the neural network solver is any one of a GRU-based solver, a Koopman operator evolution solver, or a Neural ODE solver. Specifically, the neural network solver is preferably a GRU-based solver; the structure of the GRU-based solver is as follows... Figure 2 As shown in (c), the solver updates its hidden variables using encoded input data passed through gated cyclic unit (GRU) cells.

[0028] In this embodiment of the invention, both the encoder and decoder employ a multilayer perceptron structure. The multilayer perceptron structure consists of two fully connected layers. The first fully connected layer has a ReLU activation function; the second fully connected layer has a linear function.

[0029] Both the encoder and decoder employ a two fully connected layer design, which facilitates near-linear transformation between the encoded input and output.

[0030] In this embodiment of the invention, the dimension of the hidden state vector is 8, and the number of hidden state vectors is 128. The dimension of the hidden state represents an upper bound on the independent variables required to describe the system. The value of this parameter is significantly affected by the type of model used. For P2D models, the required dimension ranges from approximately 10³ to... This depends on the level of detail in the mesh. In the context of a 1D battery model, this range typically extends from 10² to 10³. Simpler SPMe models usually require around 5 or 6 states. The most basic equivalent circuit model requires only 1 or 2 states. For the sake of analogy with theoretical models, SPMe is chosen here. This invention uses a hidden state dimension of 8 as a typical value, while also taking into account the possibility of higher-order minima.

[0031] Example 2 This invention provides a model training method, the specific steps of which are basically the same as those of the model training method provided in Embodiment 1, the only difference being that the constructed neural network-based lithium battery evaluation model is different. In this embodiment, the neural network solver of the neural network-based lithium battery evaluation model (referred to as Koopman) is a Koopman operator evolution solver, and the structure of the Koopman operator evolution solver is as follows. Figure 2 As shown in (d), the Koopman operator is dynamically updated by a GRU unit. This operator then manages the transformation between the encoded input and output.

[0032] Example 3 This invention provides a model training method, the specific steps of which are basically the same as those of the model training method provided in Embodiment 1, the only difference being that the constructed neural network-based lithium battery evaluation model is different. In this embodiment, the neural network solver of the neural network-based lithium battery evaluation model (hereinafter referred to as Neural ODE) is a Neural ODE solver. The structure of the Neural ODE solver is as follows: Figure 2 As shown in (e), the solver uses the fourth-order Runge-Kutta method to update the hidden state, relying on a neural network to learn the differential rules that control the underlying ODE system.

[0033] Example 4 like Figure 3 and Figure 4 As shown, this embodiment of the invention provides a lithium battery evaluation method based on neural networks, including the following steps: Acquire the state variable data of the lithium battery under test; The state variable data of the lithium battery to be tested is input into the target lithium battery evaluation model, and the lithium battery state variable evaluation data is output to evaluate the internal state of the lithium battery to be tested; wherein, the target lithium battery evaluation model is obtained based on the model training method of Example 3; the steps of the above method are described in detail below: (1) Calculation of time window for fusion frame: Determine the length of the time window used for multi-frame fusion, and set a reasonable window size according to the actual operating characteristics of lithium battery; (2) Calculation time and step size setting: Based on the lithium battery evaluation requirements, set the total calculation time and calculation step size (Δt). (3) Initial time setting: Determine the starting time point for calculation. ; (4) Initial internal state setting: Set the initial internal state for the neural network model; (5) Input / output tensor initialization: Initialize the input tensor and output tensor The initial value; (6) Loading conditions and historical output fusion: The current loading conditions (such as battery pack status, environmental parameters, etc.) are merged with the historical output of multi-frame fusion as input; (7) Input tensor encoding: Encode the input tensor; (8) Iterative calculation of dynamic process solver: Iterative calculation of state H is performed based on neural network iterator f; an implicit function model describing the dynamic characteristics of lithium battery (i.e. lithium battery evaluation model) is constructed, and the slope of the implicit function is calculated based on the fourth-order Runge-Kutta method; (9) Hidden state decoding: Decode the iterated hidden state into physical quantities for output; (10) Output physical quantity timing: Generate timing data of key physical quantities of lithium battery (such as voltage, current, temperature, etc.); (11) Output internal state timing: Record and output the timing changes of the internal state of the model; (12) Termination condition judgment: Determine whether the preset calculation termination condition has been met. If it has been met, stop the above evaluation process.

[0034] Example 5 This invention provides a lithium battery evaluation device, specifically comprising: The data acquisition module is used to acquire the state variable data of the lithium battery under test; The evaluation module is used to input the state variable data of the lithium battery to be tested into the target lithium battery evaluation model and output the estimated state variable data of the lithium battery to be tested, thereby evaluating the internal state of the lithium battery to be tested; wherein, the target lithium battery evaluation model is obtained based on the model training method of Example 3.

[0035] Example 6 This invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the neural network-based lithium battery evaluation method described above.

[0036] Example 7 This invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the neural network-based lithium battery evaluation method described above.

[0037] Comparative Example 1 This comparative example provides a lithium battery evaluation model based on a neural network (RNN1 for short). Its structure is basically the same as the model structure provided in Example 1, the only difference being that the neural network solver is a simple neural network, the structure of which is as follows: Figure 2 As shown in (a), the solver uses a single fully connected layer to directly map the encoded input to the output, without considering the inheritance of hidden states.

[0038] Comparative Example 2 This comparative example provides a lithium battery evaluation model based on a neural network (RNN2 for short). Its structure is basically the same as the model provided in Example 1, except that the neural network solver is a two-layer neural network, and the structure of the two-layer neural network is as follows: Figure 2 As shown in (b) of the example, the solver introduces an additional fully connected layer based on Comparative Example 1.

[0039] The following simulation experiments compare the performance of the neural network-based lithium battery evaluation models provided in Examples 1-3 of this invention with those provided in Comparative Examples 1 and 2. Training data was generated using the PyBaMM v21.10 simulation platform and divided into training and validation sets at a 9:1 ratio. The battery parameters used were NMC cathode (LiNi). 0.8 Mn 0.1 Co 0.1The system uses O2 as the base, graphite anode, and LiPF6 / EC-DMC electrolyte. Input conditions include constant current charge / discharge (10 rates from 0.05C to 1.5C) and a custom drive cycle (6 random current curves). Output parameters cover 16 physical quantities, including voltage, SOC, positive and negative electrode potentials, interface overpotential, average / surface lithium concentration, and electrolyte potential. The sampling interval is set to 1 second, generating 16 curves (each for 6000 seconds). A 30-step sliding window is used to construct 60,000 training samples. The model is implemented based on the TensorFlow framework, specifically using a GRU-based solver with 8 hidden units. The Neural ODE solver employs a fourth-order Runge-Kutta integrator with a time step h=1.0. Five models were trained and validated using training and validation sets respectively. The training results and prediction performance of each model are shown below. Figure 5 As shown.

[0040] pass Figure 5 As can be clearly seen in (a), all five model architectures converged within 100 training epochs, demonstrating that the mean squared error (MSE) was lower than that of the others. .pass Figure 5 As can be clearly seen in (b), there is a significant positive correlation between model complexity and prediction accuracy. The models provided in Example 2 (Koopman) and Example 3 (Neural ODE) perform exceptionally well, followed by the model provided in Example 1 (GRU), while the models provided in Comparative Examples 1 and 2 have the largest errors. Figure 5 As can be seen from (ce), the application of the model provided in Example 3 for generating 1000-second predictions has proven to be highly consistent with conventional ODE solutions for key internal variables, including electrode potential, interface overvoltage, and ion concentration distribution. The robustness of the five models was compared by introducing Gaussian noise into the input, and the results are as follows... Figure 5 As shown in (f) in the diagram. (Through) Figure 5 (f) shows that Comparative Example 1 and Comparative Example 2 models have the largest error increases, while Example 1 and Example 2 models have the smallest error increases. Example 3 model is relatively sensitive to errors due to its obvious slope stacking structure.

[0041] It is worth noting that as the prediction range increases, error accumulation becomes observable, such as... Figure 5As shown in (e) of the diagram. Error accumulation in long sequence forecasting is a common problem in time series forecasting models. Typically, GRU or Long Short-Term Memory (LSTM) models can achieve an optimal forecast length of approximately 20-30 steps. Therefore, "long sequence" here refers to a forecast step length much larger than the typical value, such as 1000 steps. This phenomenon can be attributed to gradient vanishing or exploding after long sequence iterations. A comparative analysis was performed using a designed validation set to evaluate the average error between 100-step and 1000-step forecasts, and the results are shown in the diagram. Figure 6 As shown. (Through) Figure 6 As can be seen in (a), the average error of 1000-step prediction remains relatively low at around 0.06. However, compared to 100-step prediction, the error increases significantly. To mitigate error accumulation in long sequences, in addition to some common methods such as gradient clipping and parameter normalization, this invention also designs three other correction methods and compares them: 1) increasing the number of hidden states; 2) indirectly constraining physical quantities; 3) increasing the length of the training sequence.

[0042] Adding extra hidden states has proven to be an effective strategy to mitigate error accumulation. In fact, estimates show that using 128 or more states can ensure that the cumulative error remains within acceptable limits above 1000 steps. However, increasing the number of hidden states leads to an exponential increase in the number of system parameters. Figure 6 As shown in (b), it provides a comprehensive overview of the correlation between model parameters and prediction accuracy, enhancing the model of Comparative Example 1 to the models provided in Examples 2 and 3, resulting in a significant improvement in prediction accuracy even with minimal additional parameter input. Simply increasing the number of units also works, but to a relatively small degree. Therefore, it can be considered a backup option.

[0043] Indirect variable modification methods have also been shown to improve the computational accuracy of models. For example... Figure 6 The SOC variables shown in (c) are numerically integrated through the charge conservation relationship, and the SOC is updated during the input reconstruction process. Figure 6 As shown in (d), the model accuracy of Example 3 after physical variable correction is significantly improved. The physical variable correction method specifically involves continuously modifying the input using a known variable (e.g., obtaining the State of Charge (SOC) through a known algorithm such as ampere-hour integration, and replacing the estimated SOC with the next input, injecting external prior information to improve the overall state estimation accuracy). This enhancement helps illustrate the model's ability to identify the interrelationships between latent variables; that is, the correction of a single variable simultaneously leads to the enhancement of other variables. This method is limited and requires consideration of additional physical assumptions in practical applications. Figure 6As shown in (e), increasing the length of the training sequence has limited effect on improving model performance. As the number of training sequences increases, the model's computational error initially decreases, then increases, reaching its optimal performance around 30-40 steps. It's important to note that insufficient sequence length may hinder the model's ability to capture long-term features, while excessively long sequences can lead to vanishing or exploding gradients. Sequences that are too long or too short can also impair the model's ability to learn relationships between latent variables, rendering constraint methods ineffective. Extended sequence lengths also hinder the efficiency of the training process. Therefore, around 30 steps is the most suitable action sequence.

[0044] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

Claims

1. A model training method, characterized in that, include: Acquire historical data of lithium battery state variables within a time window; A lithium battery evaluation model based on neural networks was built using the TensorFlow framework. The target lithium battery evaluation model is obtained by training a neural network-based lithium battery evaluation model using historical data of lithium battery state variables. The target lithium battery evaluation model is used to evaluate the internal state of the lithium battery under test based on the state variable data of the lithium battery under test.

2. The method of claim 1, wherein, The neural network-based lithium battery evaluation model includes an encoder-neural network solver-decoder architecture; The encoder is used to encode the input historical data of lithium battery state variables; The neural network solver, located between the encoder and the decoder, is used to merge the encoded historical data of the lithium battery state variables with the hidden state vector of the previous time step, and to calculate the hidden state vector of the current time step. The decoder is used to decode the hidden state vector calculated at the current time step and output lithium battery state variable evaluation data.

3. The method of claim 1, wherein, The formula for calculating the hidden state vector at the current time step is: ； where H t is the lithium battery state vector at time t, H t-1 is the lithium battery state vector at time t-1, and x t is the loading condition vector.

4. The method of claim 1, wherein, The state variables of a lithium battery include the initial state and boundary conditions; The initial state includes battery voltage, battery state of charge, positive / negative electrode potential, positive / negative electrode interface overpotential, average lithium concentration of positive / negative electrode particles, lithium concentration on the surface of positive / negative electrode particles, and electrolyte potential. Boundary conditions include lithium battery current and temperature.

5. The method of claim 1, wherein, The neural network solver can be any one of the following: GRU-based solver, Koopman operator evolution solver, or Neural ODE solver.

6. The method according to claim 1, characterized in that, Both the encoder and decoder employ a multilayer perceptron architecture.

7. The method according to claim 6, characterized in that, The multilayer perceptron structure consists of two fully connected layers; The first fully connected layer has a ReLU activation function; the second fully connected layer has a linear function.

8. The method according to claim 1, characterized in that, The dimension of the hidden state vector is 8, and the number of hidden state vectors is 128.

9. A lithium battery evaluation method, characterized in that, Includes the following steps: Acquire the state variable data of the lithium battery under test; The state variable data of the lithium battery to be tested is input into the target lithium battery evaluation model, and the lithium battery state variable evaluation data is output to evaluate the internal state of the lithium battery to be tested; wherein, the target lithium battery evaluation model is obtained based on the model training method described in any one of claims 1 to 8.

10. A lithium battery evaluation device, characterized in that, include: The data acquisition module is used to acquire the state variable data of the lithium battery under test; An evaluation module is used to input the state variable data of the lithium battery to be tested into a target lithium battery evaluation model and output the estimated state variable data of the lithium battery to be tested, thereby evaluating the internal state of the lithium battery to be tested; wherein the target lithium battery evaluation model is obtained based on the model training method described in any one of claims 1 to 8.

Images