Method and system for predicting maximum temperature of battery

Abstract

The invention provides a battery maximum temperature prediction method and system, and belongs to the field of battery safety. The method comprises the following steps: collecting time sequence data and key static parameters of a battery through a needling test; constructing a temperature-time polynomial curve model based on the temperature time sequence data and obtaining a basic prediction value; inputting the time series data, the key static parameters and the basic predicted value into a dual-model collaborative neural network model, extracting deep spatial-temporal features through a CNN-LSTM feature extraction sub-network, and calculating the correction of the basic predicted value through a BP neural network; and finally, adding the basic predicted value and the correction value, and outputting a final maximum temperature predicted value of the battery. According to the method, the physical robustness of a traditional polynomial model is combined with the nonlinear high-precision fitting capability of deep learning, so that the technical problem that the temperature prediction precision and the model generalization capability are difficult to consider at the same time in a battery acupuncture scene with small samples and high nonlinear characteristics is effectively solved.

Description

Technical Field

[0001] This invention belongs to the field of battery safety technology, and in particular relates to a method and system for predicting the maximum temperature of a battery. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] As the global new energy industry upgrades towards higher density and higher safety, the safety performance of power batteries, as the core energy carrier, directly determines the operational risks of the entire vehicle. Internal short circuits are the primary cause of battery thermal runaway. When the positive and negative electrodes of a battery come into direct contact due to external forces such as puncture or compression, a large current discharge will occur instantaneously, leading to combustion and explosion. Therefore, simulating internal short circuit scenarios through nail penetration tests and accurately predicting the battery's maximum temperature during this process has become a core technological support for assessing battery safety levels, optimizing thermal management design, and formulating early warning strategies.

[0004] However, current technologies for predicting the maximum temperature of power batteries face three major bottlenecks, making it difficult to meet the accuracy and generalization requirements of high-safety scenarios: First, traditional physical modeling methods suffer from poor adaptability and high complexity. Existing physical models based on the heat conduction equation require precise input of dozens of parameters such as specific heat capacity, thermal conductivity, and reaction heat of each component inside the battery. However, after a short circuit caused by a needle puncture, the internal structure of the battery is destroyed, and the parameters change dynamically and are difficult to measure in real time. At the same time, such models cannot effectively characterize the nonlinear coupling process between the local high temperature at the short circuit point and the overall heat dissipation. Batteries with different capacities and different cathode materials need to be remodeled, resulting in extremely high engineering application costs.

[0005] Secondly, single data-driven models suffer from limitations in both accuracy and generalization. On the one hand, while polynomial curve fitting models are simple to construct and can quickly establish a basic correlation between temperature and time, they can only capture global trends and cannot identify the temperature abrupt changes at the moment of a short circuit (and the subtle differences in the long-term stabilization phase). On the other hand, single neural network models, such as BP and LSTM, can fit nonlinear relationships, but they have significant shortcomings. LSTM models are insufficient in capturing local abrupt changes and are prone to prediction lag due to temperature jumps at the moment of a short circuit; CNN models can extract local features but cannot establish long-term dependencies on temperature changes; and both types of models are highly dependent on the distribution of training data and have weak generalization ability for battery specifications or needle penetration parameters that were not included in the training. Summary of the Invention

[0006] To overcome the shortcomings of the prior art, this invention provides a method and system for predicting the maximum temperature of a battery. The method collects the temperature, time series, and key parameters of the battery's positive electrode through a needle penetration test. After preprocessing, a temperature-time polynomial curve model is constructed to obtain the basic predicted value. Accurate correction is achieved through a dual-model collaborative neural network composed of a CNN-LSTM feature extraction subnetwork and an attention mechanism-enhanced BP correction subnetwork. Finally, the predicted result of the maximum battery temperature is output through dual-model collaborative inference.

[0007] To achieve the above objectives, one or more embodiments of the present invention provide the following technical solutions: The first aspect of this invention provides a method for predicting the maximum temperature of a battery; A method for predicting the maximum temperature of a battery includes: Data acquisition steps: Collect time-series data and key static parameters of the battery during the needle penetration test; Basic prediction steps: Based on the time series data, construct a temperature-time curve model and obtain basic prediction values; Residual correction step: Input the time series data, key static parameters and the basic predicted value into the trained dual-model collaborative neural network model, and output the correction amount of the basic predicted value; Collaborative output step: Add the base prediction value to the correction amount to output the final predicted maximum battery temperature value.

[0008] As a further technical solution, the time-series data includes temperature time series and voltage time series, and the key static parameters include battery parameters and nail penetration test condition parameters.

[0009] As a further technical solution, based on the time-series data, a temperature-time curve model is constructed, and basic predicted values ​​are obtained, including: The battery cathode temperature time series in the time series data is preprocessed, and curve fitting is performed using the least squares method to construct a fourth-order polynomial function with time as the independent variable and temperature as the dependent variable as the temperature-time curve model; its functional form is:

[0010] Where t is the time elapsed since the start of the acupuncture test, and a, b, c, d, e are the model coefficients obtained through fitting; The goodness of fit of the temperature-time curve model; The basic predicted temperature value is obtained by calculating the maximum value of the temperature-time curve model within the defined domain.

[0011] As a further technical solution, the dual-model collaborative neural network includes: The feature extraction subnetwork is used to perform deep feature extraction on the time series data and output a deep spatiotemporal feature vector. The residual prediction subnetwork is used to receive the deep spatiotemporal feature vector, the key static parameters and the basic prediction value, and calculate the correction amount based on them.

[0012] As a further technical solution, the feature extraction sub-network is a CNN-LSTM hybrid network, which first extracts local features of the time series data through a one-dimensional convolutional neural network, and then captures long-term dependencies in the feature sequence after convolution through a long short-term memory network.

[0013] As a further technical solution, the residual prediction subnetwork is an attention-enhanced backpropagation neural network; the input layer of the residual prediction subnetwork is connected to the output layer of the feature extraction subnetwork to receive the deep spatiotemporal feature vector; the input layer of the residual prediction subnetwork also receives the key static parameters and the basic prediction values.

[0014] As a further technical solution, the residual prediction subnetwork includes an attention layer after the input layer, which is used to adaptively weight the concatenated fusion feature vector. The fusion feature vector is composed of the deep spatiotemporal feature vector, the key static parameters and the basic prediction values.

[0015] A second aspect of the present invention provides a battery maximum temperature prediction system.

[0016] A battery maximum temperature prediction system includes: The data acquisition module is configured to collect the timing data and key static parameters of the battery during the needle penetration test. The basic prediction module is configured to: construct a temperature-time curve model based on the time series data and obtain basic prediction values; The residual correction module is configured to: input the time series data, key static parameters and the basic prediction value into the trained dual-model collaborative neural network model, and output the correction amount of the basic prediction value; The collaborative output module is configured to add the basic prediction value to the correction amount and output the final predicted value of the battery's maximum temperature.

[0017] A third aspect of the present invention provides a computer-readable storage medium having a program stored thereon that, when executed by a processor, implements the steps of a battery maximum temperature prediction method as described in the first aspect of the present invention.

[0018] A fourth aspect of the present invention provides an electronic device, including a memory, a processor, and a program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of a battery maximum temperature prediction method as described in the first aspect of the present invention.

[0019] The above one or more technical solutions have the following beneficial effects: (1) The CNN-LSTM feature extraction subnetwork used in this invention can extract local mutation features and long-term dependencies from time-series data, achieving a deep and comprehensive characterization of the dynamic process of thermal runaway. The attention mechanism is used to adaptively weight the multi-source fusion features, enabling the model to dynamically focus on the most critical information for prediction. This targeted network structure design ensures the maximization of data value and is the core guarantee for achieving high-precision prediction. By introducing key static parameters as input to the neural network, the model can understand and learn the differences in thermal runaway behavior of different batteries under different test conditions.

[0020] (2) This invention fully leverages the advantages of both models by employing a dual-model collaborative architecture: the multinomial model provides a stable prediction baseline that conforms to physical laws, ensuring the rationality of the method's output under extreme conditions and laying the foundation for the model's robustness; the neural network model focuses on learning complex nonlinear residuals, accurately correcting the systematic errors of the baseline model, thereby achieving a leap in prediction accuracy. This collaborative mechanism overcomes the inherent limitations of a single model. The neural network does not need to learn a complete and highly complex temperature mapping relationship from scratch, but only needs to learn a relatively simple "correction amount" mapping. This greatly reduces the learning difficulty of the model and the demand for data, effectively mitigating the risk of overfitting. This allows the invention to achieve excellent performance even under limited data conditions, possessing extremely high engineering application value and promotional feasibility.

[0021] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0022] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0023] Figure 1 This is a flowchart of the method in the first embodiment.

[0024] Figure 2 This is a system structure diagram of the second embodiment. Detailed Implementation

[0025] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0026] It should be noted that the terminology used herein is for the purpose of describing particular implementations only and is not intended to limit the exemplary implementations of the present invention.

[0027] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0028] Example 1 This embodiment discloses a method for predicting the maximum temperature of a battery. It involves collecting time-series data and key static parameters of the battery through a nail penetration test; constructing a temperature-time polynomial curve model based on the temperature time-series data and obtaining a basic predicted value; inputting the time-series data, key static parameters, and basic predicted value into a dual-model collaborative neural network model. This model extracts deep spatiotemporal features through a CNN-LSTM feature extraction subnetwork, and then calculates the correction amount of the basic predicted value through an attention-enhanced BP correction subnetwork; finally, the basic predicted value and the correction amount are added together to output the final predicted maximum battery temperature. This invention effectively solves the technical challenge of balancing temperature prediction accuracy and model generalization ability in battery nail penetration scenarios with small sample sizes and high nonlinear characteristics.

[0029] Specifically, such as Figure 1 As shown, a method for predicting the maximum temperature of a battery includes: Step S1: Collect timing data and key static parameters of the battery during the needle penetration test; Step S2: Based on the time series data, construct a temperature-time curve model and obtain basic predicted values; Step S3: Input the time series data, key static parameters and the basic prediction value into the trained dual-model collaborative neural network model, and output the correction amount of the basic prediction value; Step S4: Add the basic predicted value to the correction amount to output the final predicted value of the battery's maximum temperature.

[0030] Specifically, the above steps also include the following: Step S1: Collect the timing data and key static parameters of the battery during the needle penetration test.

[0031] The needle penetration test process includes selecting a battery with intact shape and full charge, fixing it on the test bench of the extrusion needle penetration machine, and using a data acquisition instrument to fix the temperature wire to the positive terminal of the battery; according to the standard requirements, a full needle penetration test is performed on the large surface of the battery, and the test parameters meet the following: the needle penetration test uses a high temperature resistant steel needle with a diameter of 6mm (the needle tip cone angle is 60°, the steel needle surface is smooth, rust-free, free of oxide layer and oil stains), at a speed of 25mm / s, perpendicular to the center of the large surface of the battery, and the steel needle remains in the battery after penetrating.

[0032] During the needle penetration test, time-series data and key static parameters of the battery were collected. The time-series data included the time series of the battery's positive electrode temperature and the battery voltage, collected during the needle penetration test. The key static parameters included the battery's rated capacity, initial state of charge (SOC) before the test, the diameter of the steel needle used for the needle penetration, and the needle penetration speed.

[0033] Step S2: Based on the time series data, construct a temperature-time curve model and obtain basic predicted values.

[0034] The battery positive electrode temperature time series in the time series data is preprocessed by data cleaning, including identifying and processing sudden spikes in data points that are obviously inconsistent with physical laws due to momentary sensor malfunction or electromagnetic interference. For the few missing data points caused by brief interruptions in the acquisition equipment, linear interpolation or the average of the data before and after the missing data points is used to fill in the gaps and ensure the continuity of the time series.

[0035] Furthermore, effective data segments from the start of heating to the peak temperature are extracted from the preprocessed temperature time series; data segments from the moment of needle puncture to the point before the temperature begins to decline significantly after reaching its peak are also extracted as effective modeling data. Specifically, the moment when the temperature first exceeds a certain threshold of the baseline temperature is taken as the starting point, and the moment corresponding to the global highest temperature point in the entire series is taken as the ending point. This operation aims to focus on the core heating phase of thermal runaway and eliminate interference from subsequent cooling or unstable fluctuations on model fitting.

[0036] The least squares method was used to fit the effective data segment, constructing a fourth-order polynomial temperature-time curve model with time as the independent variable and temperature as the dependent variable; its functional form is:

[0037] Where t is the time elapsed since the start of the acupuncture test, and a, b, c, d, e are the model coefficients obtained through fitting.

[0038] Calculate the goodness-of-fit R² of the temperature-time curve model and ensure it is not less than 0.9. Calculate the maximum value of the function by solving for the zeros of the first derivative of the fourth-order polynomial function in its domain, and use this maximum value as the basic predicted value. .

[0039] Step S3: Input the time series data, key static parameters and the basic predicted value into the trained dual-model collaborative neural network model, and output the correction amount of the basic predicted value.

[0040] The dual-model collaborative neural network includes a feature extraction subnetwork and a residual prediction subnetwork. The feature extraction subnetwork is used to perform deep feature extraction on the time-series data and output a deep spatiotemporal feature vector.

[0041] Specifically, the feature extraction subnetwork is a CNN-LSTM hybrid network, which processes the preprocessed temperature time series. and voltage time series Perform format adaptation and convert. A tensor of dimension, where, To represent the total number of time steps, the adapted temperature and voltage time series are input into a one-dimensional convolutional neural network for convolution calculation. The convolution kernel slides along the time dimension, and the data in the t-th step window is multiplied and summed element by element with the kernel weights, as shown below:

[0042] in, The summation result; S is the kernel size, i.e., the time window length; C is the number of input channels; It is a single convolution kernel; This is a bias term.

[0043] The corresponding feature maps are obtained by iterating through all time steps. And the feature maps are stitched together as The high-level feature sequence is obtained. In this process, the kernel is optimized through weight iteration to accurately capture local patterns such as short-term drastic temperature rise and voltage drop inflection points, while filtering instantaneous noise.

[0044] Subsequently, the output of the one-dimensional convolutional neural network (CNN) The high-level feature sequences are input into the LSTM network sequentially over time steps. The LSTM network dynamically controls the transmission and forgetting of feature information through gating mechanisms such as input gates, forget gates, and output gates, capturing long-term contextual dependencies throughout the entire time range. Specifically, let the LSTM hidden layer dimension be H, and the input at step t be the t-th time-step vector of the feature sequence. Hidden state is Cell state is The forget gate determines to discard irrelevant historical information from the cell's state, such as stable temperature data from non-critical phases, calculated as follows:

[0045] in, It is the sigmoid activation function. , This is the weight matrix.

[0046] The key features of the current step are preserved through the input gate, which is calculated as follows:

[0047]

[0048]

[0049] in, Let t be the output vector of the input gate at time step t; The input weight matrix; For the hidden weight matrix; The high-level feature sequence output by a one-dimensional convolutional network contains local features at step t. Let be the hidden state vector of the LSTM at time step t-1; , The bias vector for the input gate; This represents the candidate cell state vector; for function; , These are the input feature weight matrix and the hidden state weight matrix for the candidate cell state, respectively.

[0050] The output gate is calculated as follows:

[0051]

[0052] in, Let be the output vector of the output gate at time step t; , , These are the input feature weight matrix of the output gate, the hidden state weight matrix, and the bias vector of the output gate, respectively. Let be the final hidden state vector of the LSTM at time step t.

[0053] By iterating step-by-step, LSTM can capture long-term dependencies, such as the effect of the initial temperature rise rate on the later temperature. If the temperature rise rate is fast in a certain period, LSTM will convey this information through cell state and associate it with the prediction of the subsequent temperature peak.

[0054] Furthermore, global pooling is performed on the hidden states of the LSTM at all time steps to obtain a deep feature vector that combines detailed features such as local mutations and inflection points extracted by the one-dimensional convolutional network with long-term context dependencies captured by the LSTM. This provides comprehensive spatiotemporal feature support for the subsequent residual prediction subnetwork.

[0055] Furthermore, the residual prediction subnetwork is used to receive the deep spatiotemporal feature vector, the key static parameters, and the basic predicted value, and calculate the correction amount based on them. The residual prediction subnetwork is an attention-enhanced backpropagation (BP) neural network, which receives the deep feature vector from the feature extraction subnetwork. The key static parameter vector S characterizing the test conditions includes parameters such as battery rated capacity, initial SOC, steel needle diameter, and needle penetration speed, as well as basic predicted values ​​from the multinomial model. The information from these three different sources is concatenated to form a comprehensive fused feature vector. This vector simultaneously contains the dynamic spatiotemporal characteristics of battery thermal runaway, static test conditions, and baseline predictions, providing comprehensive information support for subsequent accurate corrections.

[0056] The residual prediction subnetwork includes an attention layer after the input layer to adaptively weight the concatenated fused feature vector. Specifically, the fused feature vector... Feed it into an attention layer, through a small trainable neural network, for... Each feature dimension in the dataset is assigned a weight score. Specifically, a weight score is calculated for each feature dimension. Input a single-hidden-layer neural network and generate intermediate features through linear transformation and tanh activation function. Secondly, regarding intermediate features Perform linear mapping and Softmax activation, and calculate weight scores:

[0057] in, For mapping vectors, For bias. The weight score represents the weight value; a larger weight value indicates that the corresponding feature dimension is more important for the prediction of the correction amount.

[0058] The third step is to assign weights. Features of original fusion Element-wise multiplication yields the weighted eigenvector:

[0059] in, This is the weighted fusion feature vector, which enhances key features and suppresses redundant information, thereby improving the prediction accuracy of the subsequent BP network for the correction amount.

[0060] Then the weighted features The data is fed into a subsequent backpropagation (BP) neural network consisting of several fully connected layers and nonlinear activation functions to achieve the mapping from high-dimensional features to scalar correction values. First, the data is processed through connected layers... A linear transformation is performed, followed by ReLU activation to filter out invalid features and introduce non-linearity. Higher-order features are further extracted by repeating the linear transformation and ReLU activation in the hidden layer. Finally, the output layer outputs a correction value by performing a linear transformation on the final hidden layer features. .

[0061] Step S4: Add the basic predicted value to the correction amount to output the final predicted value of the battery's maximum temperature.

[0062] base forecast value Correction amount from the dual-model collaborative neural network computation module Perform a summation calculation and calculate the result. The final predicted maximum battery temperature will be output as shown in the following formula:

[0063] in, This is the predicted final maximum battery temperature value.

[0064] Example 2 This embodiment discloses a battery maximum temperature prediction system; like Figure 2 As shown, a battery maximum temperature prediction system includes: The data acquisition module is configured to collect the timing data and key static parameters of the battery during the needle penetration test. The basic prediction module is configured to: construct a temperature-time curve model based on the time series data and obtain basic prediction values; The residual correction module is configured to: input the time series data, key static parameters and the basic prediction value into the trained dual-model collaborative neural network model, and output the correction amount of the basic prediction value; The collaborative output module is configured to add the basic prediction value to the correction amount and output the final predicted value of the battery's maximum temperature.

[0065] Example 3 The purpose of this embodiment is to provide a computer-readable storage medium.

[0066] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in a battery maximum temperature prediction method as described in Example 1.

[0067] Example 4 The purpose of this embodiment is to provide an electronic device.

[0068] An electronic device includes a memory, a processor, and a program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps in a battery maximum temperature prediction method as described in Example 1.

[0069] The steps and methods involved in the apparatuses of Embodiments 2, 3, and 4 above correspond to those in Embodiment 1. For specific implementation details, please refer to the relevant description section of Embodiment 1. The term "computer-readable storage medium" should be understood as a single medium or multiple media including one or more instruction sets; it should also be understood as including any medium capable of storing, encoding, or carrying an instruction set for execution by a processor and enabling the processor to perform any of the methods in this invention.

[0070] Those skilled in the art will understand that the modules or steps of the present invention described above can be implemented using general-purpose computer devices. Optionally, they can be implemented using computer-executable program code, thereby allowing them to be stored in a storage device for execution by a computer device, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. The present invention is not limited to any particular combination of hardware and software.

[0071] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A battery maximum temperature prediction method characterized by, The method comprises the following steps: a data collection step: collecting time series data and key static parameters of the battery during the needle test; a basic prediction step: based on the time series data, a temperature-time curve model is constructed, and a basic prediction value is obtained; a residual correction step: the time series data, key static parameters and the basic prediction value are input into the trained double-model collaborative neural network model, and a correction amount of the basic prediction value is output; a collaborative output step: the basic prediction value and the correction amount are added, and a final battery maximum temperature prediction value is output.

2. The battery maximum temperature prediction method of claim 1, wherein, The time series data includes temperature time series and voltage time series, and the key static parameters include battery parameters and needle test condition parameters.

3. The battery maximum temperature prediction method of claim 1, wherein, Based on the time series data, a temperature-time curve model is constructed, and a basic prediction value is obtained, including: The battery positive electrode temperature time series in the time series data is preprocessed, and the least squares method is used for curve fitting to construct a quartic polynomial function with time as the independent variable and temperature as the dependent variable as the temperature-time curve model; the function form is: where t is the time elapsed since the beginning of the needle test, a, b, c, d, e are model coefficients obtained by fitting; is the goodness of fit of the temperature-time curve model; The basic prediction value of the temperature is obtained by calculating the maximum value of the temperature-time curve model in the definition domain.

4. The battery maximum temperature prediction method of claim 1, wherein, The double-model collaborative neural network comprises: a feature extraction subnetwork for deep feature extraction of the time series data, outputting a deep spatio-temporal feature vector; a residual prediction subnetwork for receiving the deep spatio-temporal feature vector, the key static parameters and the basic prediction value, and calculating the correction amount based thereon.

5. A battery maximum temperature prediction method according to claim 4, characterized in that, The feature extraction subnetwork is a CNN-LSTM hybrid network, which first extracts local features of the time series data through a one-dimensional convolutional neural network, and then captures long-term dependencies in the feature sequence after convolution processing through a long short-term memory network.

6. The battery maximum temperature prediction method of claim 4, wherein, The residual prediction subnetwork is an attention mechanism enhanced back propagation neural network; the input layer of the residual prediction subnetwork is connected to the output layer of the feature extraction subnetwork, and is used to receive the deep spatio-temporal feature vector; the input layer of the residual prediction subnetwork also receives the key static parameters and the basic prediction value.

7. A battery maximum temperature prediction method according to claim 6, wherein The residual prediction subnetwork comprises an attention layer after the input layer, which is used for adaptive weighting of the fused feature vector after splicing, and the fused feature vector is spliced from the deep spatio-temporal feature vector, the key static parameters and the basic prediction value.

8. A battery maximum temperature prediction system characterized by, The method comprises the following steps: a data collection module configured to collect time series data and key static parameters of the battery during the needle test; a basic prediction module configured to construct a temperature-time curve model based on the time series data, and obtain a basic prediction value; a residual correction module configured to input the time series data, key static parameters and the basic prediction value into the trained double-model collaborative neural network model, and output a correction amount of the basic prediction value; a collaborative output module configured to add the basic prediction value and the correction amount, and output a final battery maximum temperature prediction value.

9. A computer-readable storage medium having stored thereon a program, characterized in that, The program is executed by the processor to realize the steps in the battery maximum temperature prediction method according to any one of claims 1-7.

10. An electronic device comprising a memory, a processor, and a program stored on the memory and executable on the processor, characterized in that, The processor implements the steps in the battery maximum temperature prediction method according to any one of claims 1-7 when executing the program.

Images