CN-122021773-A - Battery thermal runaway time prediction model training method, prediction device and prediction method thereof

CN122021773ACN 122021773 ACN122021773 ACN 122021773ACN-122021773-A

Abstract

The invention belongs to the technical field of battery safety detection, and discloses a battery thermal runaway time prediction model training method, a prediction device and a prediction method thereof. The invention discloses a battery thermal runaway time prediction model training method which comprises the following steps of collecting battery related multi-source time sequence monitoring data and integrating the data into a structured input sequence according to a unified time reference, establishing a staged fusion neural network model, wherein a TCN module adopts a stacked residual block and an expansion causal convolution to extract multi-scale local time sequence characteristics, a transducer encoder captures the overall long-term dependence of the sequence through a multi-head self-attention mechanism, the two are fused in a progressive structure depth, and after training and optimizing the model by utilizing a training set, the trained model is deployed to a battery management system to realize real-time online accurate prediction of the thermal runaway residual time. According to the invention, the accuracy and timeliness of the thermal runaway early warning are obviously improved through the complementary advantages of the local feature extraction and the global associated modeling.

Inventors

HUANG YAJUN
LI WEI
WANG TIANTIAN
WANG ZHIRONG
WANG JUNLING
Shen Xiongxuan
LIU JIALONG

Assignees

南京工业大学

Dates

Publication Date: 20260512
Application Date: 20260129

Claims (9)

1. A battery thermal runaway time prediction model training method, characterized in that the model training method comprises the following steps: Step S1, multi-source time sequence monitoring data of a battery under the condition of operation or heat abuse is obtained, wherein the multi-source time sequence monitoring data at least comprises one or more of temperature, temperature rise rate, voltage, gas release parameters, quality change or expansion force, a multi-dimensional time sequence input sequence is constructed according to a uniform time scale, an original data set formed by the multi-dimensional time sequence input sequence and a label is divided into a training set, a verification set and a test set, and the data dimension influence is removed by utilizing normal distribution normalization; Step S2, inputting the multidimensional time sequence input sequence preprocessed in the step S1 into a time sequence convolution network TCN module, modeling local dynamics of the multidimensional time sequence input sequence, and outputting a risk candidate feature sequence representing a thermal runaway evolution trend, wherein the time sequence convolution network TCN module sequentially comprises a plurality of residual blocks, and feature extraction is carried out through the plurality of residual blocks, each residual block at least comprises two layers of expansion causal convolution, an activation function and 1X 1 convolution, and the expansion rate of the expansion causal convolution is gradually increased in different residual blocks according to a preset rule of geometric progression increment; Step S3, taking the risk candidate feature sequence obtained in the step S2 as input of a transducer encoder, wherein the transducer encoder consists of a plurality of encoder blocks which are stacked in sequence, and each encoder block sequentially comprises a multi-head attention module, a first Dropout and normalization module, a feed-forward network module and a second Dropout and normalization module; Step S4, carrying out global average pooling processing on the global risk representation obtained in the step S3 in a time dimension to eliminate the influence of instantaneous fluctuation on a prediction result and obtain an overall risk state representation; Step S5, training the staged fusion neural network model in the steps S2-S4 by using a training set, verifying the trained model by using verification set data, and checking whether the model is trained to an optimal state; And S6, inputting the test set into the trained staged fusion neural network model to obtain a prediction result, and evaluating and analyzing the prediction precision by applying an average absolute error, a mean square error, a root mean square error and a decision coefficient R2.
2. The method for training a thermal runaway time prediction model of a battery according to claim 1, wherein the battery is a lithium iron phosphate battery, and the thermal runaway data, i.e., the multi-source time series monitoring data in step S1, includes a front surface temperature, a front surface temperature rise rate, a rear surface temperature rise rate, a left surface temperature rise rate, a right surface temperature rise rate, a terminal voltage, a mass loss, a hydrogen concentration, a carbon monoxide concentration, a methane concentration, an ethylene concentration, and an expansion force of the battery.
3. The method for training a thermal runaway time prediction model of a battery according to claim 1, wherein in step S1, the segmentation ratio of the training set, the verification set and the test set is 7:1:2.
4. The method for training a thermal runaway time prediction model of a battery according to claim 1, wherein in step S1, the specific calculation formula of the normal distribution normalization is: ; Wherein X * represents normalized data; sample data is represented, μ represents the mean; representing standard deviation.
5. The method for training a thermal runaway time prediction model of a battery according to claim 1, wherein in the step S2, the expansion rates of the expansion causal convolution in the plurality of residual blocks in the time series convolution network module are 1,2 and 4, respectively.
6. The method of claim 1, wherein in the step S3, the self-attention mechanism module predicts the encoded data set using the query vector Q, the key vector K, and the value vector V, and a learnable weight matrix corresponding to the change, and the output generated by the module is: ; where softmax (i) V represents the weighted sum, d k represents the dimensions of the query vector Q and the key vector K, and K T represents the transpose of the K vector.
7. The method for training a battery thermal runaway time prediction model according to claim 1, wherein the training of the staged fusion neural network model in the step S5 includes the steps of optimizing the staged fusion neural network model using an Adam optimizer, iterating the model a plurality of times using the training set and updating parameters by a specified batch size, and evaluating the staged fusion neural network model by a verification set after each iteration.
8. A battery thermal runaway time prediction device is used for executing the battery thermal runaway time prediction model training method according to claim 1, and comprises a data acquisition module, a data preprocessing module, a model training module and a prediction module, wherein the data acquisition module is used for synchronously acquiring multi-source time sequence monitoring data, the data preprocessing module is used for executing normalization processing on the multi-source time sequence monitoring data and constructing characteristics and labels, normalizing the data to eliminate numerical value differences among the data, dividing a data set into a training set, a verification set and a test set, the model training module is used for training the preprocessed training set as input of a staged fusion neural network model to obtain a thermal runaway prediction model, and the prediction module is used for inputting test set data into the trained staged fusion neural network model to obtain a prediction result.
9. A battery thermal runaway time prediction method is characterized in that the optimal staged fusion neural network model trained in the battery thermal runaway time prediction model training method is deployed and applied, specifically, the battery thermal runaway time prediction method is embedded into a monitoring unit of a target battery system, real-time sensing data after pretreatment is received, residual thermal runaway time is continuously calculated and output, thermal runaway early warning information is generated according to residual time prediction values, and the thermal runaway early warning information is used for guiding safety management or emergency response of the battery system.

Description

Battery thermal runaway time prediction model training method, prediction device and prediction method thereof Technical Field The invention belongs to the technical field of battery safety detection, and particularly relates to a battery thermal runaway time prediction model training method, a prediction device and a prediction method thereof. Background Lithium iron phosphate batteries have become key energy devices in the fields of electric automobiles, large-scale energy storage and the like due to high safety and long cycle life. However, under extreme abuse conditions, such lithium iron phosphate batteries may still undergo thermal runaway, resulting in serious safety accidents. At present, thermal runaway early warning mainly relies on threshold monitoring of single parameters such as voltage or temperature. The method is simple but has limitations. Thermal runaway is the result of complex coupling of multiple physical fields, and single parameters are easily disturbed, resulting in low early warning accuracy and insufficient reliability. Meanwhile, the fixed threshold method is difficult to describe a time sequence evolution rule among parameters, and early warning cannot be provided. Some electrochemical-thermal coupling models are proposed, but the electrochemical-thermal coupling models are complex in calculation and depend on accurate internal parameters, so that the requirements of actual online application are difficult to meet. In recent years, deep learning models represented by a Recurrent Neural Network (RNN) and a long-short-term memory network (LSTM) have been attempted to be used for battery state of health prediction due to their strong timing characteristic capturing ability. However, when these sequence models deal with long sequence prediction, there is a general problem that the gradient disappears or explodes, and the recursive structure thereof causes low calculation efficiency, and accumulated errors can spread with the increase of the prediction step length, so as to affect the stability of long-term prediction. In addition, the thermal runaway process is usually accompanied by severe and nonlinear mutation of data, and the prediction model is required to have not only excellent long-term memory capability, but also high sensitivity to key mutation point characteristics and strong parallel computing efficiency, which is difficult to be compatible with the traditional time sequence model. Therefore, it is necessary to construct a prediction model capable of deeply fusing the time sequence local mode and the global multi-parameter association, so as to early warn the thermal runaway of the lithium phosphate battery more accurately and earlier, and improve the reliability of safety management. Disclosure of Invention The invention aims to provide a battery thermal runaway time prediction model training method, and the constructed prediction model is a mixed frame model of a fusion time sequence convolutional network (TCN) and a transducer, so that the time sequence dynamic characteristics and the global dependency relationship of a thermal runaway precursor are deeply excavated, the problems of low accuracy and poor timeliness of the traditional early warning method are solved, and the early and reliable prediction of the thermal runaway is realized. The invention adopts the following technical scheme: In a first aspect, the present invention provides a method for training a thermal runaway time prediction model of a battery, where the thermal runaway time prediction model of a battery is a thermal runaway time prediction model of a battery based on a staged fusion neural network, and specifically the method for training a model includes the following steps: Step S1, multi-source time sequence monitoring data of a battery under the condition of operation or heat abuse are obtained, for example, the battery is heated in a closed combustion chamber experiment platform in a mode of applying 600W constant power to heat the battery to be tested to trigger thermal runaway, the multi-source time sequence monitoring data at least comprise one or more of temperature, temperature rise rate, voltage, gas release parameters, quality change or expansion force, a multi-dimensional time sequence input sequence is built according to a uniform time scale, an original data set formed by the multi-dimensional time sequence input sequence and a label is divided into a training set, a verification set and a test set, and the influence of data dimension is removed by normalization of normal distribution. The multi-dimensional time sequence input sequence is composed of a plurality of monitoring parameters aligned with time stamps, the label is the residual time from each time stamp to the occurrence point of thermal runaway, and the value of the residual time is determined by the difference value between the thermal runaway starting judgment time stamp and the current time stamp. Step S2, inputting the