Battery thermodynamic parameter determination method independent of calorimetric data

Abstract

The invention discloses a battery thermodynamic parameter determination method independent of calorimetric data, and relates to the technical field of battery systems, and the method comprises the following specific steps: (1) constructing a sampling matrix of input parameters of a thermal runaway model; (2) establishing a battery thermal runaway numerical calculation model and executing high-throughput simulation to generate a pre-training data set; (3) constructing a multi-stage neural network model and carrying out pre-training; (4) carrying out transfer learning fine tuning by using small-batch experimental data to realize the adaptation of the model in the experimental data set; and (5) substituting the identified thermodynamic parameters into the thermal runaway model for verification and optimization. According to the method, the thermodynamic parameters can be directly identified based on the temperature data measured by the battery thermal abuse experiment, the complex calorimetric experiment process is avoided, and the test cost is reduced.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of battery systems, and particularly relates to a battery thermal dynamics parameter determination method independent of calorimetric data. BACKGROUND

[0002] As an important part of modern energy systems, secondary batteries represented by lithium-ion batteries have been widely used in electric vehicles, energy storage power stations, and portable electronic devices. However, under abuse or extreme working conditions, a series of intense exothermic reactions may occur inside the battery, leading to rapid temperature rise and triggering thermal runaway. Thermal runaway not only causes single battery fire or explosion, but also may cause chain reaction at the module or system level, triggering thermal runaway propagation and causing serious safety hazards.

[0003] In order to reveal the mechanism of battery thermal runaway and improve the intrinsic safety of battery systems, it is essential to accurately obtain the thermal dynamics parameters inside the battery. Thermal dynamics parameters are key indicators that describe the characteristics of exothermic reactions inside the battery, including activation energy, pre-exponential factor, and reaction enthalpy. These parameters are also the core input of the battery thermal runaway prediction model, which determines the reaction rate and heat release intensity during thermal runaway evolution. If the thermal dynamics parameters are not accurate, the model will have prediction deviation for the temperature evolution process, and it is difficult to truly reflect the occurrence and propagation behavior of thermal runaway. Therefore, it is of great significance to establish an efficient, accurate, and universal battery thermal dynamics parameter determination method to improve the reliability and accuracy of battery thermal safety prediction.

[0004] Currently, the determination of battery thermal dynamics parameters mainly relies on calorimetric experiments, such as accelerated calorimeter and differential scanning calorimeter tests. By fitting the calorimetric data, the activation energy and pre-exponential factor of side reactions can be obtained, thereby providing input for thermal runaway models (as described in the literature Appl Energy. 2023; 336: 120695.; eTransportation. 2022; 12: 100157.). However, this method has obvious limitations: first, the calorimetric experiment process is complex, long and costly, and the experimental operation has safety risks; second, the parameter identification process relies on researchers' manual division and curve fitting of temperature rise stages, which is highly subjective and prone to errors; third, the calorimetric results are usually only applicable to specific chemical systems or state of charge, and are difficult to generalize to different batteries.

[0005] To solve the above problems, the application provides a battery thermal dynamics parameter determination method independent of calorimetric data. SUMMARY

[0006] The application aims to provide a battery thermal dynamics parameter determination method independent of calorimetric data to solve the above problems in the prior art.

[0007] The technical scheme is as follows: a battery thermal dynamics parameter determination method independent of calorimetric data, comprising the following steps: S1, constructing a sampling matrix of thermal runaway model input parameters; S2, establishing a battery thermal runaway numerical calculation model and performing high-throughput simulation to generate a pre-training data set; S3, constructing a multi-stage neural network model and pre-training; S4, performing transfer learning fine-tuning using small batch experimental data to adapt the model to the experimental data set; and S5, substituting the identified thermal dynamics parameters into the thermal runaway model for verification and optimization.

[0008] Further, the specific steps of S1 include:

[0009] S1.1, sampling the boundary condition parameters and thermal dynamics parameters by a Latin hypercube sampling method.

[0010] The boundary condition parameters include: (1) a heater heat flux density Q h ; (2) a battery surface convection heat transfer coefficient h; (3) a battery surface emissivity ε; and (4) an equivalent heat transfer coefficient κ of the battery and the surrounding solid structure. The thermal dynamics parameters include: (5) a trigger temperature T onset,i of a side reaction, wherein i represents the number of side reactions, and preferably i=5; (6) a trigger temperature T TR of thermal runaway; (7) a peak temperature T peak of thermal runaway; and (8) a reaction rate slope r' i of the side reaction at the trigger temperature.

[0011] The steps of the Latin hypercube sampling method include: (1) determining the value ranges of the first boundary condition parameter Q h , the second boundary condition parameter h, the third boundary condition parameter ε, the fourth boundary condition parameter κ, the first thermal dynamics parameter T peak , and the second thermal dynamics parameter r' i according to the statistical results of actual thermal runaway test data; and (2) determining the value range of the thermal runaway trigger temperature T TR , and randomly sampling in the range to obtain a sampling trigger temperature T TR,s; (3) divide the temperature interval of the trigger temperature of the first side reaction to the sampling trigger temperature T TR,s into five continuous subintervals, in turn, [T1, T2], [T2, T3], [T3, T4], [T4, T5], [T5, T TR,s ]; (4) according to the length of the subinterval [T i , T i+1 ], correct the value range of r' i, so that the corrected R' i is inversely proportional to the interval length; (5) divide the parameter value range and the temperature subinterval into N subintervals according to equal probability; (6) randomly select a sample point in each subinterval; (7) randomly combine all parameter sample points, so that each sample point is used only once, forming N groups of sample matrices with uniform distribution characteristics.

[0012] S1.2, convert the trigger temperature and the reaction rate slope into the pre-exponential factor, the activation energy and the reaction enthalpy. The specific steps include:

[0013] (1) solve the reaction rate calculation relationship, convert the linear parameter space with T onset,i and r' i as variables into the exponential parameter space with A i and Ea i as variables:

[0014]

[0015]

[0016] wherein r onset represents the critical reaction rate of reaction triggering, the value is 0.01; A i represents the pre-exponential factor; Ea i represents the activation energy; R represents the molar gas constant; T onset,i is the trigger temperature of the side reaction, and r' i is the reaction rate slope, obtained from sampling in S1;

[0017] (2) calculate the enthalpy ΔH i of different reaction stages according to the trigger temperature and the peak temperature of the side reaction:

[0018]

[0019] wherein m is the battery mass, and C p is the specific heat capacity.

[0020] S1.3, organize all sampling parameters into a thermal runaway model input matrix, and save it in CSV format for subsequent simulation calling.

[0021] Further, the specific steps of S2 include:

[0022] S2.1, establishing a battery thermal runaway numerical calculation model. The basic theory and establishment process include:

[0023] In the thermal runaway process, the heat released by the electrochemical reaction will continuously increase the battery temperature. Considering the calculation efficiency and accuracy of high-throughput simulation, a thermal runaway model based on thermal resistance network is used to describe this process:

[0024]

[0025]

[0026]

[0027] T c is the battery temperature, T a is the ambient temperature; t is time; Q TR is the heat release of the side reaction in the thermal runaway process; Qh represents the heater heat flux density; c i is the dimensionless concentration of active materials; R i is the thermal resistance between adjacent nodes, and the specific calculation method is listed in Table 1.

[0028] S2.2, calling the sampling matrix file described in S1.3 to build a batch simulation task.

[0029] S2.3, based on the parallel computing framework, the batch simulation task is solved by high-throughput to obtain the battery thermal runaway temperature curve.

[0030] S2.4, each temperature curve is automatically processed, and the sample that does not trigger thermal runaway is removed; the sample that does not trigger thermal runaway is defined as the sample with temperature rise rate dT / dt <1℃ / s at any time.

[0031] S2.5, generate a pre-training data set. The input features of the pre-training data set are the battery temperature time series, and the time interval is 1s; the output label is the boundary condition parameter and the thermal dynamic parameter, including: the heater heat flux density Q h , the battery surface convection heat transfer coefficient h, the battery surface emissivity ε, the equivalent heat transfer coefficient κ of the battery and the surrounding solid structure, the trigger temperature T onset,i of the side reaction, and the reaction rate slope r' i.

[0032] Table 1 Calculation formula of thermal resistance between battery nodes

[0033]

[0034] Further, the specific steps of S3 include:

[0035] S3.1, divide the simulation obtained temperature time series into three sub-sequences of heating stage, reaction stage and cooling stage, and adjust the temperature sequence of each stage to a uniform length by intercepting or zero padding to ensure the consistency of input data dimension.

[0036] The division of the heating stage and the reaction stage is based on: c =69℃;

[0037] The division of the reaction stage and the cooling stage is based on: c =T peak ;

[0038] S3.2, maximum-minimum standardization processing is performed on all input and output data.

[0039] S3.3, for the key parameters of battery temperature evolution in different stages, a multi-stage neural network model is established:

[0040] (1) Cooling stage model: local features of temperature time series are extracted using convolutional neural network (CNN), time series dependence between time series is captured using long short-term memory network (LSTM), and mapping relationship between temperature change features and boundary heat transfer parameters (including convective heat transfer coefficient h, surface emissivity ε and equivalent heat transfer coefficient κ of battery and solid structure) is established through full connection layer;

[0041] (2) Heating stage model: the boundary condition parameters h, ε and κ obtained in the previous step are part of the input layer, which are input into the CNN-LSTM model together with the temperature time series of the heating stage to establish the mapping relationship between temperature change features and external heating heat flux density Q h ;

[0042] (3) Reaction stage model: based on the calculated boundary condition parameters, the heat release rate inside the battery is calculated through the energy conservation equation, and the CNN-LSTM model is used to establish the mapping relationship between the heat release rate and the reaction kinetics parameters (including the trigger temperature T onset,i and the reaction rate slope r' i of different stage side reactions) by taking the time series as input.

[0043] S3.4, use Adam optimizer for training, and introduce adaptive learning rate to dynamically adjust the training step; after training, the output model weight is used as the initial parameter of transfer learning.

[0044] Further, the specific steps of S4 include:

[0045] S4.1, obtain the thermal runaway temperature-time sequence data of the battery under actual working conditions through literature research, local heating test and other means, and obtain the corresponding reference thermal dynamics parameters by manually fitting the experimental curve; the experimental data form a small batch of experimental data set after screening, and preferably, the number of samples is 5% of the number of pre-training set samples.

[0046] S4.2, smooth filtering is performed on the experimental temperature curve, noise points and invalid segments are removed, and the temperature time sequence is divided into three subsequences of heating stage, reaction stage and cooling stage.

[0047] S4.3, the multi-stage neural network model obtained in step S3 is used as the initial model of transfer learning, the CNN layer parameters are frozen, only the LSTM layer, the fully connected layer and the output layer are trainable, so as to retain the time sequence feature expression ability learned in the simulation domain and reduce the overfitting risk of small sample training.

[0048] S4.4, the experimental temperature data are used as the model input, and the back propagation algorithm is used to fine-tune the trainable layer parameters; during the training process, the Adam optimizer is used, and the early stopping mechanism and adaptive learning rate adjustment strategy are introduced to avoid overfitting and ensure the convergence stability.

[0049] Further, the specific steps of S5 include:

[0050] S5.1, T onset,i predicted by the transfer learning neural network model in S4 is compared with the experimental T i . i . i .

[0051] S5.2, the boundary condition parameters and the thermal dynamics parameters are input into the thermal runaway numerical calculation model in S2 to obtain the battery transient temperature evolution curve.

[0052] S5.3, the thermal runaway trigger temperature T onset (dT / dt>1℃ / s), the thermal runaway trigger time t onset , the peak temperature T peak and other characteristic parameters in the original temperature curve and the simulation temperature curve are extracted.

[0053] S5.4, the root mean square error (RMSE), the mean absolute percentage error (MAPE) and the determination coefficient (R 2 ) between the experimental and simulation characteristic parameters are calculated, which are used for comprehensive evaluation of the accuracy of the model prediction.

[0054] S5.5, if the error exceeds the preset threshold, return to S3 to adjust the training weight or return to S4 to fine-tune the network parameters, until the error between the model prediction result and the experimental result meets the requirements.

[0055] S5.6, save the optimized model weight and the set of thermokinetic parameters, form a general thermokinetic parameter calculation model suitable for different battery systems, state of charge and abuse working conditions.

[0056] The advantages of the present application compared with the prior art are: (1) the present application breaks through the dependence of traditional thermokinetic parameter identification on calorimetric experiments, and can realize the identification of thermokinetic parameters based on the temperature evolution data obtained in thermal abuse experiments, avoiding the problems of long cycle, high cost, complex operation and safety risk of calorimetric experiments; (2) the present application can quickly determine the key thermokinetic parameters (activation energy, pre-exponential factor, reaction enthalpy, etc.) within seconds, greatly improving the efficiency of thermal runaway model development; (3) the efficient mapping of simulation domain and experimental domain is realized through the transfer learning strategy, so that the model can complete rapid adaptation under the condition of small sample experimental data, significantly enhancing the universality and cross-chemical system portability of the thermokinetic parameter identification method; (4) the present application constructs a complete "parameter sampling-high-throughput simulation-model training-transfer learning-error feedback optimization" closed loop process, which can automatically output high-precision battery thermokinetic parameters, and provides an efficient, low-cost and universal solution for battery thermal runaway modeling, digital twin and safety management, which has significant engineering application and popularization value. BRIEF DESCRIPTION OF DRAWINGS

[0057] Figure 1 is an operation flowchart of a battery thermokinetic parameter determination method independent of calorimetric data according to the present application; Figure 1

[0058] Figure 2 is an operation flowchart of generating a pre-training data set by high-throughput simulation in the present application; Figure 2

[0059] Figure 3 is a schematic diagram of battery nodes and thermal resistance in the embodiment of the present application; Figure 3

[0060] Figure 4 is a calculation flowchart of a multi-stage neural network model in the present application; Figure 4

[0061] Figure 5 is a comparison diagram of temperature curves obtained by experimental measurement, pre-training model prediction and transfer learning model prediction in the embodiment of the present application. Figure 5

[0062] Figure 6 is a comparison diagram of actual and predicted thermal runaway characteristic parameters in the embodiment of the present application. Figure 6 DETAILED DESCRIPTION

[0063] ​​​​​​The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the protection scope of the present invention.

[0064] Example

[0065] Taking a prismatic battery as an example, based on the transient temperature curve of thermal runaway under thermal abuse conditions, the thermodynamic parameters of the battery are determined, and the predicted temperature curve is compared with the actual measured curve to provide a comprehensive and detailed description of the invention. This method is not limited to prismatic batteries but is applicable to lithium-ion batteries of different shapes and chemical systems. (See attached...) Figure 1 As shown, the invention mainly consists of the following steps: S1, constructing a sampling matrix of input parameters for the thermal runaway model; S2, establishing a numerical calculation model for battery thermal runaway and performing high-throughput simulation to generate a pre-training dataset; S3, constructing a multi-stage neural network model and performing pre-training; S4, using small-batch experimental data for transfer learning fine-tuning to achieve model adaptation to the experimental dataset; S5, substituting the identified thermodynamic parameters into the thermal runaway model for verification and optimization.

[0066] S1, Construct the sampling matrix of the input parameters for the thermal runaway model. Specific steps include:

[0067] S1.1, the boundary condition parameters and thermodynamic parameters are sampled separately using the Latin hypercube sampling method.

[0068] The boundary condition parameters include: (1) heater heat flux density Q h (2) Convective heat transfer coefficient h of the battery surface; (3) Emissivity ε of the battery surface; (4) Equivalent heat transfer coefficient κ between the battery and the surrounding solid structure. The thermodynamic parameters include: (5) Triggering temperature T of the side reaction. onset,i i represents the number of side reactions, preferably i=5; (6) the trigger temperature T for thermal runaway TR (7) Peak temperature of thermal runaway T peak (8) The slope of the reaction rate of the side reaction at the trigger temperature, r'i.

[0069] The steps of the Latin hypercube sampling method include: (1) determining the first boundary condition parameter Q based on the statistical results of the actual thermal runaway test data. h Second boundary condition parameter h, third boundary condition parameter ε, fourth boundary condition parameter κ, first thermodynamic parameter T peak (2) Determine the range of values ​​for the second thermodynamic parameter r'i; (3) Determine the thermal runaway trigger temperature T.TR The range of values ​​is determined, and random sampling is performed within this range to obtain the sampling trigger temperature T. TR,s (3) Adjust the trigger temperature of the first side reaction to the sampling trigger temperature T. TR,s The temperature range is divided into five consecutive sub-ranges, namely [T1, T2], [T2, T3], [T3, T4], [T4, T5], [T5, T... TR,s (4) Based on the subinterval [T] i , T i+1 The length of ] is used to correct the range of values ​​of r' i, so that the corrected R' i is inversely proportional to the length of the interval; (5) the range of parameter values ​​and temperature sub-intervals are divided into N sub-intervals with equal probability; (6) a sample point is randomly selected in each sub-interval; (7) the sample points of all parameters are randomly combined so that each sample point is used only once, forming N sets of sample matrices with uniform distribution characteristics.

[0070] S1.2 converts the trigger temperature and reaction rate slope into pre-exponential factor, activation energy, and reaction enthalpy.

[0071] S1.3 organizes all sampled parameters into a thermal runaway model input matrix and saves it in CSV format for subsequent simulation calls.

[0072] S2. Establish a numerical model for battery thermal runaway and perform high-throughput simulations to generate a pre-trained dataset. (See attached image) Figure 2 As shown, the specific steps of S2 include:

[0073] S2.1, Establish a numerical calculation model for battery thermal runaway; the connection method of the thermal resistance is shown in the attached figure. Figure 3 As shown.

[0074] S2.2, call the sampling matrix file described in S1.3 to construct a batch simulation task.

[0075] S2.3, Based on the parallel computing framework, the batch simulation task is solved in high throughput to obtain the battery thermal runaway temperature curve.

[0076] S2.4 Automated processing is performed on each temperature curve to remove samples that have not triggered thermal runaway; the sample that has not triggered thermal runaway is defined as a sample whose temperature rise rate dT / dt at any time is less than 1℃ / s.

[0077] S2.5, generate the pre-trained dataset.

[0078] S3, Construct a multi-stage neural network model and perform pre-training. Specific steps include:

[0079] S3.1, the simulation obtained temperature time series is divided into heating stage, reaction stage and cooling stage three sub-sequences, the temperature sequence of each stage is adjusted to the uniform length by intercepting or zero padding, so as to ensure the consistency of input data dimension.

[0080] The division of the heating stage and the reaction stage is based on T c =69℃;

[0081] The division of the reaction stage and the cooling stage is based on T c =T peak ;

[0082] S3.2, all input and output data are subjected to maximum-minimum standardization processing.

[0083] S3.3, for the key parameters of battery temperature evolution in different stages, neural network models are respectively established, as shown in the accompanying drawings: Figure 4

[0084] (1) cooling stage model: local features of temperature time series are extracted by using convolutional neural network (CNN), time series dependence between time series is captured by using long short-term memory network (LSTM), and the mapping relationship between temperature change characteristics and boundary heat transfer parameters (including convective heat transfer coefficient h, surface emissivity ε and equivalent heat transfer coefficient κ of battery and solid structure) is established through full connection layer;

[0085] (2) heating stage model: the boundary condition parameters h, ε and κ obtained in the previous step are part of the input layer, which are input into the CNN-LSTM model together with the temperature time series of the heating stage, so as to establish the mapping relationship between temperature change characteristics and external heating heat flux density Q h ;

[0086] (3) reaction stage model: based on the calculated boundary condition parameters, the heat release rate in the battery is calculated through the energy conservation equation, and the CNN-LSTM model is used to establish the mapping relationship between the heat release rate and the reaction kinetics parameters (including the trigger temperature T onset,i and the reaction rate slope r' i of different stage side reactions).

[0087] S3.4, Adam optimizer is used for training, and adaptive learning rate is introduced to dynamically adjust the training step; after training, the output model weight is used as the initial parameter of transfer learning.

[0088] S5, the identified thermal kinetics parameters are substituted into the thermal runaway model for verification and optimization, and the specific steps include:

[0089] S5.1, T​onset,i and r' i is converted to A i , Ea i and ΔH i .

[0090] S5.2, input the boundary condition parameters and the thermokinetic parameters into the thermal runaway numerical calculation model in S2 to obtain the battery transient temperature evolution curve.

[0091] To verify the applicability of the proposed battery thermokinetic parameter determination method independent of calorimetric data in different chemical system batteries, four typical batteries were selected for parameter identification and verification analysis. The selected batteries include: LiNi0.8Co0. 15 Al0. 05 O2 / graphite system lithium ion battery (data source: Appl Therm Eng. 2021; 189: 116661.), LiNi0.8Co0.1Mn0.1O2 / graphite system lithium ion battery (data source: Appl Therm Eng. 2021; 192: 116949.), LiFePO4 / soft carbon system solid-state battery (data source: J Energy Storage. 2025; 105: 114701.), and sodium ion battery of layered oxide / soft carbon system (data source: Process Saf Environ Protect. 2025; 193: 842-55). The thermal runaway test data in them are all from published literature. The Figure 5 The temperature curve comparison results of experimental measurement, pre-trained model prediction and transfer learning model prediction parameter calculation are shown respectively. From the figure, it can be seen that the three models have consistent change trend in the initial heating temperature rise stage, while differences appear in the temperature drop stage after exhaust. This is mainly because the pre-trained model can effectively capture the heat transfer flux Q h on the surface of the battery, but cannot fully consider the transient heat release change caused by exhaust. After transfer learning fine-tuning, the model can automatically correct Q h according to the experimental data, so as to accurately predict the trigger time and temperature evolution characteristics of thermal runaway. Quantitative analysis results show that the relative error of thermal runaway trigger time predicted by the pre-trained model is 8.23%, 51.60%, 58.76% and 9.73% respectively; after transfer learning, the relative error is reduced to 3.23%, 7.96%, 8.37% and 0.96% respectively. This shows that the transfer learning model significantly improves the accuracy of the prediction of the trigger time and transient temperature of thermal runaway of different types of batteries. In terms of peak temperature prediction, both the pre-trained model and the transfer learning model are in good consistency with the experimental results. This is because the reaction enthalpy ΔH iThe calculation is based on the energy conservation equation, rather than a model-dependent free fitting, thereby reducing the dependence of the model on empirical parameters and ensuring the physical consistency of the results. Overall, the experimental temperature curve is highly consistent with the calculated temperature curve based on the transfer learning model prediction parameters, proving the effectiveness of the invention.

[0092] S5.3, extract the thermal runaway trigger temperature T in the original temperature curve and the simulation temperature curve onset (DT / dt > 1℃ / s), thermal runaway trigger time t onset , peak temperature T peak and other characteristic parameters.

[0093] S5.4, calculate the root mean square error (RMSE), mean absolute percentage error (MAPE) and determination coefficient (R 2 ) between the experimental and simulation characteristic parameters, for comprehensive evaluation of the accuracy of the model prediction.

[0094] The Figure 6 shows the comparison of actual and predicted thermal runaway characteristic parameters. The results show that the predicted values of most samples are highly consistent with the actual values, and the scatter points are concentrated near the diagonal. Taking the thermal runaway trigger time as an example, the RMSE, MAPE and R 2 of the predicted value and the actual value are 64.96, 0.067 and 0.884, respectively, indicating that the model has high precision in predicting the characteristic parameters of the main thermal runaway stage.

[0095] S5.5, if the error exceeds the preset threshold, return to S3 to adjust the training weight or return to S4 to fine-tune the network parameters, until the error between the model prediction result and the experimental result meets the requirements.

[0096] S5.6, save the optimized model weight and thermal dynamics parameter set, and form a general thermal dynamics parameter calculation model applicable to different battery systems, state of charge and abuse conditions.

[0097] In summary, the present application proposes a battery thermal dynamics parameter determination method independent of calorimetric data. This method combines high-throughput numerical simulation with multi-stage neural networks, and only requires temperature evolution data from battery thermal abuse experiments to accurately identify thermal dynamics parameters. This method does not require complex calorimetric experimental processes, significantly improving parameter identification efficiency and safety, while having good universality and scalability, and being applicable to different types, structures and conditions of battery systems, providing an efficient and universal new technical means for battery thermal runaway mechanism research, thermal safety prediction and digital twin model construction.

Claims

1. A method for determining battery thermodynamic parameters without relying on calorimetric data, characterized in that, Includes the following steps: S1, construct the sampling matrix of input parameters for the thermal runaway model; S2, Establish a numerical calculation model for battery thermal runaway and perform high-throughput simulation to generate a pre-training dataset; S3, construct a multi-stage neural network model and perform pre-training; S4 utilizes small batches of experimental data for transfer learning fine-tuning to adapt the model to the experimental dataset. S5. The identified thermodynamic parameters are substituted into the thermal runaway model for verification and optimization.

2. The method for determining battery thermodynamic parameters without relying on calorimetric data according to claim 1, characterized in that, Step S1 includes the following operations: S1.1, the boundary condition parameters and thermodynamic parameters are sampled separately using the Latin hypercube sampling method; S1.2 converts the trigger temperature and reaction rate slope into pre-exponential factor, activation energy, and reaction enthalpy; S1.3 organizes all sampled parameters into a thermal runaway model input matrix and saves it in CSV format for subsequent simulation calls.

3. The method for determining battery thermodynamic parameters without relying on calorimetric data according to claim 1, characterized in that, The boundary condition parameters mentioned in step S1.1 include: (1) heater heat flux density Q h (2) Convective heat transfer coefficient h of the battery surface; (3) Emissivity ε of the battery surface; (4) Equivalent heat transfer coefficient κ between the battery and the surrounding solid structure. The thermodynamic parameters include: (5) Triggering temperature T of the side reaction. onset,i i represents the number of side reactions; (6) the trigger temperature T for thermal runaway TR (7) Peak temperature of thermal runaway T peak (8) The slope of the reaction rate of the side reaction at the trigger temperature, r'i; The Latin hypercube sampling step described in step S1.2 includes: (1) determining the boundary condition parameter Q based on the statistical results of the actual thermal runaway test data. h h, ε, κ and thermodynamic parameter T peak (2) Determine the range of values ​​for r'i; (3) Determine the thermal runaway trigger temperature T. TR The range of values ​​is determined, and random sampling is performed within this range to obtain the sampling trigger temperature T. TR,s (3) Adjust the trigger temperature of the first side reaction to the sampling trigger temperature T. TR,s The temperature range is divided into five consecutive sub-ranges, namely [T1, T2], [T2, T3], [T3, T4], [T4, T5], [T5, T... TR,s (4) Based on the subinterval [T] i , T i+1 The length of ] is used to correct the range of values ​​of r'i so that the corrected R'i is inversely proportional to the length of the interval; (5) the range of parameter values ​​and temperature sub-intervals are divided into N sub-intervals with equal probability; (6) a sample point is randomly selected in each sub-interval; (7) the sample points of all parameters are randomly combined so that each sample point is used only once, forming N sets of sample matrices with uniform distribution characteristics.

4. The method for determining battery thermodynamic parameters without relying on calorimetric data according to claim 1, characterized in that, Step S2 includes the following operations: S2.1, Establish a numerical calculation model for battery thermal runaway; S2.2, Call the sampling matrix file described in S1.3 to construct a batch simulation task; S2.3, Based on the parallel computing framework, the batch simulation task is solved in high throughput to obtain the battery thermal runaway temperature curve; S2.4, each temperature curve is processed automatically to remove samples that have not triggered thermal runaway; the sample that has not triggered thermal runaway is defined as a sample whose temperature rise rate dT / dt at any time is less than 1℃ / s; S2.5 generates a pre-trained dataset.

5. The method for determining battery thermodynamic parameters without relying on calorimetric data according to claim 1, characterized in that, The input features of the pre-training dataset are battery temperature time series with a time interval of 1 second; the output labels are boundary condition parameters and thermodynamic parameters, including: heater heat flux density Q. h The following parameters are considered: battery surface convective heat transfer coefficient h, battery surface emissivity ε, equivalent heat transfer coefficient κ between the battery and surrounding solid structures, and the triggering temperature T of the side reaction. onset,i 、Reaction rate slope r'i.

6. The method for determining battery thermodynamic parameters without relying on calorimetric data according to claim 1, characterized in that, Step S3 includes the following operations: S3.1, the temperature time series obtained from the simulation is divided into three subsequences: heating stage, reaction stage and cooling stage. The temperature series of each stage is adjusted to a uniform length by truncation or zero padding to ensure the consistency of the input data dimensions. S3.2, perform max-min standardization on all input and output data; S3.3, establish a multi-stage neural network model for the key parameters that dominate the battery temperature evolution at different stages; S3.4 uses the Adam optimizer for training and introduces an adaptive learning rate to dynamically adjust the training step size; After training is complete, the output model weights are used as the initial parameters for transfer learning.

7. The method for determining battery thermodynamic parameters without relying on calorimetric data according to claim 1, characterized in that, The multi-stage neural network model includes: (1) Cooling stage model: The local features of the temperature time series are extracted by using a convolutional neural network (CNN), the temporal dependence between time series is captured by a long short-term memory network (LSTM), and the mapping relationship between temperature change features and boundary heat transfer parameters (including convective heat transfer coefficient h, surface emissivity ε and equivalent heat transfer coefficient κ of battery and solid structure) is established through a fully connected layer. (2) Heating stage model: The boundary condition parameters h, ε, and κ obtained in the previous step are used as part of the input layer and input together with the temperature time series of the heating stage into the CNN-LSTM model to establish the temperature change characteristics and the external heating heat flux density Q. h The mapping relationship; (3) Reaction stage model: Based on the calculated boundary condition parameters, the heat release rate inside the battery is calculated through the energy conservation equation. Using this time series as input, a CNN-LSTM model is used to establish the mapping relationship between the heat release rate and the reaction kinetic parameters (including the trigger temperature T of side reactions at different stages). onset,i and the reaction rate slope r'i).

8. The method for determining battery thermodynamic parameters without relying on calorimetric data according to claim 1, characterized in that, Step S4 includes the following operations: S4.1, Time series data of thermal runaway temperature of battery under actual working conditions are obtained through literature review, local heating test and other means. The corresponding reference thermodynamic parameters are obtained by manually fitting the experimental curve. The experimental data are then screened to form a small batch of experimental datasets. S4.2, smooth the experimental temperature curve, remove noise points and invalid segments, and divide the temperature time series into three sub-sequences: heating stage, reaction stage and cooling stage. S4.3, use the multi-stage neural network model obtained in step S3 as the initial model for transfer learning, freeze the CNN layer parameters, and retain only the LSTM layer, fully connected layer and output layer as trainable, so as to retain the temporal feature representation ability learned in the simulation domain and reduce the risk of overfitting in small sample training. S4.4, the experimental temperature data is used as the model input, and the backpropagation algorithm is used to fine-tune the parameters of the trainable layer; During training, the Adam optimizer is used, and an early stopping mechanism and adaptive learning rate adjustment strategy are introduced to avoid overfitting and ensure convergence stability.

9. The method for determining battery thermodynamic parameters without relying on calorimetric data according to claim 1, characterized in that, Step S5 includes the following operational steps: S5.1, the T predicted by the transfer learning neural network model in S4 onset,i And r' i transformed into A i Ea i and ΔH i ; S5.2, input the boundary condition parameters and thermodynamic parameters into the thermal runaway numerical calculation model in S2 to obtain the transient temperature evolution curve of the battery; S5.3 Extract the thermal runaway trigger temperature T from the original temperature curve and the simulation temperature curve. onset Thermal runaway trigger time t onset Peak temperature T peak Characteristic parameters; S5.4 Calculate the root mean square error (RMSE), mean absolute percentage error (MAPE), and coefficient of determination (R²) between experimental and simulation characteristic parameters. 2 ), used to comprehensively evaluate the accuracy of model predictions; S5.5 If the error exceeds the preset threshold, return to S3 to adjust the training weights or return to S4 to fine-tune the network parameters until the error between the model prediction result and the experimental result meets the requirements. S5.6 saves the optimized model weights and thermodynamic parameter set to form a general thermodynamic parameter calculation model applicable to different battery systems, states of charge, and abuse conditions.

Images