Method for predicting thermal runaway of vehicle based on time sequence data

Abstract

The invention provides a method for predicting vehicle thermal runaway based on time sequence data, and relates to the technical field of vehicle thermal runaway prediction. According to the method, a network model of a time sequence is constructed based on a traditional neural network, and the traditional neural network model only suitable for section data is expanded into a panel data model for new energy automobile thermal runaway; according to the method, the thermal runaway fault caused by a single data reason can be continued, and the reason causing the thermal runaway can be represented on sequential time nodes, that is, the fault representation can be continued, so that the thermal runaway of the new energy automobile can be effectively and accurately predicted. By adopting the novel neural network model disclosed by the invention, progressive characteristics of historical data of a time sequence can be fully utilized, the prediction capability of the model is mined to a greater extent, and the accuracy of the model is greatly improved through actual use prediction.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of vehicle thermal runaway prediction, in particular to a method for predicting vehicle thermal runaway based on time series data. BACKGROUND

[0002] With the rapid development of automobile technology, the optimization and prediction of vehicle thermal systems have become key factors in improving automobile performance and energy efficiency. In new energy vehicle battery systems, thermal runaway is a serious problem that can have a significant impact on the safety and stability of the vehicle.

[0003] Traditional vehicle thermal runaway prediction methods are usually based on physical models and statistical methods, but these methods often show limitations in dealing with complex nonlinear systems. Neural networks, as a powerful machine learning tool, can handle a large number of nonlinear relationships and perform well in pattern recognition, making them an ideal choice for studying vehicle thermal runaway prediction. In contrast, neural networks can predict new vehicle battery data by learning from historical fault vehicle data, with higher accuracy and robustness. By constructing a neural network model and training it with a large amount of vehicle operation-related training data, accurate prediction of vehicle thermal runaway can be achieved. This method greatly improves the accuracy of prediction.

[0004] However, existing neural network models are usually only suitable for cross-sectional data and difficult to handle time series data and panel data; this limitation makes it difficult for the model to effectively capture the continuity of thermal runaway failures in time series. Traditional methods are difficult to handle thermal runaway problems caused by multiple factors acting together, and cannot effectively identify and analyze the impact of single data on thermal runaway. Existing time series prediction methods are mostly based on the stationarity assumption, making it difficult to effectively model complex nonlinear time series data, especially when dealing with long-term dependencies; that is, traditional neural network models often ignore the periodicity and trend characteristics of data when processing time series data, resulting in deficiencies in prediction accuracy. Therefore, it is particularly important to develop a method that can accurately predict thermal runaway of new energy vehicle batteries. SUMMARY

[0005] The present application aims to at least solve one of the technical problems existing in the prior art.

[0006] To this end, the present application provides a method for predicting vehicle thermal runaway based on time series data.

[0007] The present application provides a method for predicting vehicle thermal runaway based on time series data, comprising: obtaining training data, the training data being time series data of vehicle operating parameters; The training data is data sliced, wherein the training data is data sliced according to a time sequence and a set time window size; A thermal runaway prediction model based on a neural network is established, an input of the thermal runaway prediction model is a vehicle operating parameter in single slice data, and an output is a thermal runaway probability value; the thermal runaway prediction model includes a plurality of input layers connected in sequence, wherein the vehicle operating parameter of the last time node in the set time window is input to the first layer input layer, and the vehicle operating parameters of the remaining time nodes in the set time window are sequentially input to the remaining input layers as additional input quantities according to a time sequence until all the time node data in the set window is used up; A plurality of slice data and corresponding thermal runaway probability value labels are used as sample data, and the sample data is used to train the thermal runaway prediction model; Vehicle operating data is obtained, and the vehicle operating data and the training data are used to represent the time sequence state during the operation of the vehicle and the battery pack; The vehicle operating data is input into the trained thermal runaway prediction model to perform real-time thermal runaway prediction on the battery. Since the battery or the battery pack is part of the vehicle, real-time thermal runaway prediction on the battery is also thermal runaway prediction on the vehicle.

[0008] According to the above technical scheme of the present application, the method for predicting vehicle thermal runaway based on time sequence data can also have the following additional technical features: In the above technical scheme, the set time window includes n time nodes, and the thermal runaway prediction model includes n-1 input layers; The main input quantity of the first layer input layer includes the vehicle operating parameter of the nth time node, and the vehicle operating parameter of the n-1th time node is used as an additional input quantity of the first layer input layer; Each of the 2th to n-1th input layers is defined as an intermediate input layer, the main input quantity of the i th intermediate input layer includes the output quantity of the i-1th input layer, and the vehicle operating parameter of the n-i th time node is used as an additional input quantity of the i th intermediate input layer; wherein i is greater than or equal to 2 and less than or equal to n-1.

[0009] In the above technical scheme, each input layer of the thermal runaway prediction model includes m neurons, the number of neurons in each input layer is equal to the number of feature data types in the vehicle operating parameter; in the first layer input layer, the total input quantity of each neuron is:

[0010]

[0011]

[0012]

[0013]

[0014] wherein, represents the total input of the jth neuron in the 1st input layer, ; represents the jth feature data in the vehicle operating parameters at the nth time node; represents the parameter of ; represents the jth feature data in the vehicle operating parameters at the (n-1)th time node; represents the parameter of .

[0015] In the above technical solution, in the intermediate input layer, the total input of each neuron is:

[0016]

[0017]

[0018]

[0019]

[0020] wherein, represents the total input of the jth neuron in the ith input layer; represents the output of the jth neuron in the (i-1)th input layer; represents the parameter of ; represents the jth feature data in the vehicle operating parameters at the (n-i)th time node; represents the parameter of .

[0021] In the above technical solution, each neuron in the input layer uses a sigmod activation function to generate an output, and the expression of the sigmod activation function is:

[0022] wherein, represents the output of the neuron; y represents the total input of the neuron; e represents the natural exponential.

[0023] In the technical solution, the thermal runaway prediction model further comprises a plurality of hidden layers, the plurality of hidden layers are connected in sequence and connected with the n-1th input layer, and the total input amount of neurons in the first hidden layer connected with the n-1th input layer is represented as:

[0024]

[0025]

[0026]

[0027]

[0028] wherein, represents the total input amount of the jth neuron in the hidden layer connected with the n-1th input layer; represents the output amount of the jth neuron in the n-1th input layer; represents the parameter of The total input amount of neurons in the remaining hidden layers is represented as:

[0029]

[0030]

[0031]

[0032]

[0033] wherein, represents the total input amount of the jth neuron in the x+1th hidden layer, X represents the total number of hidden layers; represents the output amount of the jth neuron in the xth hidden layer; represents the parameter of Each neuron in the hidden layer uses a sigmod activation function to generate an output amount.

[0034] In the technical solution, the vehicle operating parameters include some or all of the following: state of charge, high voltage state, operating mode, vehicle speed, cumulative mileage, total voltage inside the battery pack, total current of the battery pack, battery pack insulation value, highest single cell voltage of the battery pack, lowest single cell voltage of the battery pack, highest temperature of the battery pack, lowest temperature of the battery pack, total number of batteries, total number of temperature probes, and temperature values of each temperature probe. ​​Before model training using training data and real-time thermal runaway prediction using vehicle operation data, data standardization processing is performed on vehicle operation parameters; wherein dummy variable conversion is used for data conversion on state of charge, high voltage state and operation mode; minimum-maximum value method is used for data conversion on vehicle speed, cumulative mileage, total voltage inside battery pack, total current of battery pack and battery pack insulation value; z-score method is used for data conversion on battery pack highest single cell voltage, battery pack lowest single cell voltage, battery pack highest temperature, battery pack lowest temperature, battery total number, temperature probe total number and temperature value of each temperature probe.

[0035] In the above technical solution, the loss function of the thermal runaway prediction model adopts a cross-entropy loss function.

[0036] In the above technical solution, when training the thermal runaway prediction model using sample data, different set time window sizes are used to generate slice data for model training respectively; The prediction effect of the trained thermal runaway prediction model is evaluated; wherein the prediction effect of the model with different set time window sizes is evaluated respectively, and the set time window size of the final model is determined according to the evaluation result; When using the thermal runaway prediction model to perform real-time thermal runaway prediction on the battery, according to the determined set time window size n', the vehicle operation parameters generated by the vehicle at the current time node and the vehicle operation parameters corresponding to the previous n'-1 historical time nodes are divided into the same time window, and the vehicle operation parameters in the time window are input into the thermal runaway prediction model to perform real-time thermal runaway prediction on the battery.

[0037] In the above technical solution, the evaluation of the prediction effect of the trained thermal runaway prediction model comprises: The prediction effect of the trained model is tested using a test set, and the accuracy, precision and recall of the prediction are calculated, and the test set also includes time series data of vehicle operation parameters and corresponding thermal runaway probability value labels. The calculation method comprises:

[0038]

[0039]

[0040] wherein, Accuracy is represented by Acc; Precision is represented by Prec; Recall is represented by Rec; The number of test results in which the slice data corresponds to a positive class true thermal runaway label and the model prediction result is also a positive class test result is represented by TP; denotes the number of test results in which the real thermal runaway label corresponding to the slice data in the test result is negative and the model prediction result is also negative; denotes the number of test results in which the real thermal runaway label corresponding to the slice data in the test result is negative and the model prediction result is positive; denotes the number of test results in which the real thermal runaway label corresponding to the slice data in the test result is positive and the model prediction result is negative.

[0041] Therefore, the present application has the following beneficial effects: By adopting the novel neural network model, the progressive characteristics of the historical data of the time series can be fully utilized, the prediction capability of the model is greatly mined, and the accuracy of the model is greatly improved through actual use prediction.

[0042] Specifically, the present application constructs a network model of time series based on a traditional neural network, expands the traditional neural network model suitable for cross-section data into a panel data model for thermal runaway of new energy vehicles, and can continue thermal runaway failure caused by single data and the cause of thermal runaway will be manifested at the subsequent time nodes, that is, the fault characteristic will continue, which can effectively and accurately predict thermal runaway of new energy vehicles.

[0043] Additional aspects and advantages of the present application will become apparent from the following description with reference to the accompanying drawings. BRIEF DESCRIPTION OF DRAWINGS

[0044] The above and / or additional aspects and advantages of the present application will become apparent and more readily appreciated from the following description of the embodiments, taken in conjunction with the accompanying drawings of which: Figure 1 is a flowchart of a method for predicting vehicle thermal runaway based on time series data according to an embodiment of the present application; Figure 2 is a neural network architecture diagram of the first layer input layer in the thermal runaway prediction model according to an embodiment of the present application; Figure 3 is a neural network architecture diagram of the multi-layer input layer in the thermal runaway prediction model according to an embodiment of the present application; Figure 4 is a neural network architecture diagram of the whole thermal runaway prediction model according to an embodiment of the present application. DETAILED DESCRIPTION

[0045] In order to more clearly understand the above-mentioned purposes, features and advantages of the present application, the present application will be further described in detail below in conjunction with the drawings and specific embodiments. It should be noted that the embodiments of the present application and the features in the embodiments can be combined with each other without conflict.

[0046] In the following description, numerous specific details are set forth in order to provide a thorough understanding of the present application. However, it will be apparent to one skilled in the art that the present application can be practiced without the specific details set forth in this description. In other instances, well-known methods, procedures, components, and circuits have not been described in detail so as not to unnecessarily obscure aspects of the present application.

[0047] Some embodiments of the present application are described below with reference to Figures 1 to 4 a method for predicting thermal runaway of a vehicle based on time series data.

[0048] Some embodiments of the present application provide a method for predicting thermal runaway of a vehicle based on time series data.

[0049] As shown in Figure 1 , a first embodiment of the present application proposes a method for predicting thermal runaway of a vehicle based on time series data, comprising the following steps S1-S6. It should be noted that the order of the above steps S1-S6 is only illustrative, and those skilled in the art can adjust the specific order of each step according to the needs, and the steps in different orders can also be executed simultaneously.

[0050] S1, obtaining training data, the training data being time series data of vehicle operating parameters, that is, the training data is composed of vehicle operating parameters corresponding to a plurality of time nodes, and the training data is usually historical vehicle operating data.

[0051] In some embodiments, the vehicle operating parameters include some or all of the following: state of charge, high voltage state, operating mode, vehicle speed, cumulative mileage, total voltage inside the battery pack, total current of the battery pack, battery pack insulation value, highest single cell voltage of the battery pack, lowest single cell voltage of the battery pack, highest temperature of the battery pack, lowest temperature of the battery pack, total number of batteries, total number of temperature probes, and temperature values of each temperature probe. It can be understood that the more comprehensive the types of vehicle operating parameters are, the more characteristic information that can be captured to affect the thermal runaway of the vehicle. In addition to the above-mentioned data types, those skilled in the art can also add other types of vehicle operating parameters as needed. In this disclosure, the vehicle operating parameters are taken as an example to include all the above-mentioned types of data.

[0052] In some embodiments, step S1 further comprises performing data standardization processing on the vehicle operating parameters in the training data, so that the vehicle operating parameters are suitable for subsequent model application.

[0053] In one specific embodiment, the data standardization process comprises: The charging state, high voltage state and running mode are converted by using a dummy variable conversion method. Specifically, the dummy variable conversion is a method for converting a classification variable into a numerical form, and the state data of the charging state, high voltage state and running mode are suitable for the dummy variable conversion method. The specific method of the dummy variable conversion is well known to those skilled in the art and is not described here.

[0054] The vehicle speed, cumulative mileage, total voltage inside the battery pack, total current of the battery pack and battery pack insulation value are converted by using a minimum-maximum value method. It can be understood that the vehicle speed, cumulative mileage, total voltage inside the battery pack, total current of the battery pack and battery pack insulation value usually have a relatively fixed value range, and their distribution is relatively stable, and it is not easy to appear extreme value, so it is suitable for the minimum-maximum value method, and the expression of the method is:

[0055] wherein, is the independent variable in the minimum-maximum value method; is the minimum value in the independent variable; is the maximum value in the independent variable.

[0056] The highest cell voltage of the battery pack, the lowest cell voltage of the battery pack, the highest temperature of the battery pack, the lowest temperature of the battery pack, the total number of batteries, the total number of temperature probes and the temperature value of each temperature probe are converted by using a z-score method. It can be understood that the highest cell voltage of the battery pack, the lowest cell voltage of the battery pack, the highest temperature of the battery pack, the lowest temperature of the battery pack, the total number of batteries, the total number of temperature probes and the temperature value of each temperature probe have a relatively large and unstable value range, and are easily affected by various factors to appear abnormal values. Therefore, it is suitable for the z-score method, and the expression of the method is:

[0057] wherein, is the independent variable in the z-score method; mean represents the average value in the independent variable; represents the standard deviation in the independent variable.

[0058] It can be understood that when the vehicle running data is used for real-time thermal runaway prediction, the vehicle running data also needs to be standardized.

[0059] S2. Slice the vehicle operating parameters, wherein the vehicle operating parameters are sliced ​​according to the time series and the set time window size. The set time window size can be set to 500, 1000, 2000, 3000, etc., and the number is used to represent the number of time nodes in the window. The larger the window, the more data it includes and the more features it has, but it has a greater impact on subsequent training. In a specific embodiment, when slicing the collected vehicle operating parameters according to the time series, no distinction is made between charging segments and non-charging segments. It is understandable that in the data slices sliced ​​according to the time series, the vehicle operating parameters should be sorted according to the time series.

[0060] S3. Establish a thermal runaway prediction model based on a neural network, wherein the input of the thermal runaway prediction model is the vehicle operating parameters in a single slice data, and the output is a thermal runaway probability value; the thermal runaway prediction model includes several input layers connected in sequence, wherein the vehicle operating parameters of the last time node in the set time window are input into the first input layer, and the vehicle operating parameters of the remaining time nodes in the set time window are input into the remaining input layers in sequence according to the time series as additional inputs, until all the time node data in the set window are used up.

[0061] Specifically, the thermal runaway prediction model in this disclosure is built on a traditional neural network. The main difference is that this disclosure transforms the traditional neural network's single-layer input into a multi-layer input, strengthening the progressive relationship of time series data and the relationship between adjacent data. In other words, the thermal runaway prediction model includes not only multiple input layers, but also the hidden and output layers of traditional neural networks.

[0062] In some embodiments, if the time window is set to include n time nodes, n-1 input layers are set in the thermal runaway prediction model.

[0063] Among them, the main input of the first input layer includes the vehicle operating parameters at the nth time node, and the vehicle operating parameters at the n-1th time node serve as the additional input of the first input layer; Each input layer from the 2nd to the n-1th input layer is defined as an intermediate input layer, then the main input of the i-th intermediate input layer includes the output of the i-1th input layer, and the vehicle operating parameters at the ni-th time node are used as the additional input of the i-th intermediate input layer; where i is greater than or equal to 2 and less than or equal to n-1.

[0064] In a specific embodiment, Figure 2 As shown, each input layer of the thermal runaway prediction model includes m neurons. The number of neurons in each input layer is equal to the number of characteristic data types in the vehicle operating parameters. In the first input layer, the total input of each neuron is:

[0065]

[0066]

[0067]

[0068]

[0069] wherein, represents the total input of the jth neuron in the 1st input layer, ; represents the jth feature data in the vehicle operating parameters at the nth time node; represents parameters of the function; represents the jth feature data in the vehicle operating parameters at the (n-1)th time node; represents parameters of the function.

[0070] In one specific embodiment, as shown in Figure 3 , in the intermediate input layer, the total input of each neuron is:

[0071]

[0072]

[0073]

[0074]

[0075] wherein, represents the total input of the jth neuron in the ith input layer; represents the output of the jth neuron in the (i-1)th input layer; represents parameters of the function; represents the jth feature data in the vehicle operating parameters at the (n-i)th time node; represents parameters of the function.

[0076] In some embodiments, each neuron in the above-mentioned input layers generates an output using a sigmod activation function, and the expression of the sigmod activation function is:

[0077] wherein, represents the output of the neuron; y represents the total input of the neuron; and e represents the natural exponent.

[0078] Based on the input provided by the aforementioned network, and in accordance with the technical characteristics of traditional neural network models, a hidden layer and an output layer are added to the end of the last input layer to calculate the thermal runaway probability. This disclosure uses multiple hidden layers to fit a nonlinear function. The specific number of hidden layers can be selected based on actual needs to balance training difficulty and prediction accuracy.

[0079] In some embodiments, as Figure 4 As shown, the thermal runaway prediction model includes several hidden layers, which are sequentially connected to the n-1th input layer. The total input amount of neurons in the hidden layer connected to the n-1th input layer, that is, the first hidden layer, is expressed as:

[0080]

[0081]

[0082]

[0083]

[0084] in, represents the total input of the jth neuron in the hidden layer connected to the n-1th input layer; represents the output of the jth neuron in the n-1th input layer; express Parameters; The total input to the neurons in the remaining hidden layers is expressed as:

[0085]

[0086]

[0087]

[0088]

[0089] in, represents the total input of the jth neuron in the x+1th hidden layer, , X represents the total number of hidden layers; represents the output of the jth neural unit in the xth hidden layer; express Parameters; Each neuron in the hidden layer also uses a sigmod activation function to generate an output quantity. Finally, the output result, i.e., the thermal runaway probability value, is determined according to the output quantity of the last hidden layer.

[0090] S4, a plurality of slice data and corresponding thermal runaway probability value labels are used as sample data, the thermal runaway prediction model is trained using the sample data, and the thermal runaway prediction model parameters are updated through a back propagation algorithm.

[0091] It can be understood that the above-mentioned data slices are known to have thermal runaway or not in the corresponding time window, that is, each data slice has a corresponding true thermal runaway label, and the training set formed according to a plurality of sample data is used to train the thermal runaway prediction model.

[0092] In some embodiments, the loss function of the thermal runaway prediction model uses a cross-entropy loss function. The expression of the cross-entropy loss function is as follows:

[0093] Wherein, L represents the cross-entropy loss function, N represents the number of samples; The true label of sample g is represented by g, and the positive class is 1 and the negative class is 0. Specifically, the meanings of the positive class and the negative class can be set as needed, for example, the occurrence of thermal runaway of the vehicle represents the positive class, and the non-occurrence of thermal runaway of the vehicle represents the negative class; Pi represents the probability that sample i is predicted as a positive class.

[0094] In some embodiments, the back propagation algorithm uses an Adam gradient descent algorithm, and the learning rate is set to 0.001. The specific content of the Adam gradient descent algorithm is known to those skilled in the art, and will not be described here.

[0095] In some embodiments, the slice data generated by different set time window sizes can also be used for model training in step S4; for example, time series data with window sizes of 500, 1000, 2000 and 3000 are used for model training in turn, and the default training is 1000 epochs. Further, the prediction effects of the models under different set time windows are analyzed and compared.

[0096] S5, the prediction effect of the trained thermal runaway prediction model is evaluated.

[0097] In some embodiments, step S5 includes testing the prediction effect of the trained model using a test set, and calculating the accuracy, precision and recall of the prediction, and the test set also includes time series data of vehicle operating parameters and corresponding thermal runaway probability value labels. The calculation method includes:

[0098]

[0099]

[0100] wherein, represents the accuracy rate; represents the precision rate; represents the recall rate; represents the number of test results in which the slice data corresponds to a true hot runaway label of the positive class and the model prediction result is also a positive class in the test result; represents the number of test results in which the slice data corresponds to a true hot runaway label of the negative class and the model prediction result is also a negative class in the test result; represents the number of test results in which the slice data corresponds to a true hot runaway label of the negative class and the model prediction result is a positive class in the test result; represents the number of test results in which the slice data corresponds to a true hot runaway label of the positive class and the model prediction result is a negative class in the test result.

[0101] In some embodiments, the prediction effect of the model with different set time window sizes can be evaluated respectively, and the set time window size of the final model selection can be determined according to the evaluation result.

[0102] In one specific embodiment, the model evaluation is performed on the time series data with the set window size, i.e., 500, 1000, 2000, and 3000, and the calculated values of the accuracy rate, the precision rate, and the recall rate corresponding to each set time window are shown in Table 1.

[0103] Table 1. Model evaluation result statistics corresponding to different set time windows

[0104] It can be seen that the larger the window setting of the time series, the higher the accuracy rate, but the precision rate and the recall rate are not necessarily better with the larger window setting of the time series. In particular, the abnormal amplitude value and the sensitivity of the abnormality of a certain type of data in the hot runaway data may cause the sensitivity to be not sensitive in the large window time series. After data standardization, the abnormal characteristics of the data may be hidden, resulting in a decrease in the precision rate and the recall rate. Therefore, the set time window is selected by step S5 in the present disclosure, and the set time window size is set to 2000 time node steps in the present embodiment.

[0105] S6, obtain vehicle operation data, the vehicle operation data and the training data are used to represent the time sequence state in the operation process of the vehicle and the battery pack, that is, the vehicle operation data is also composed of vehicle operation parameters corresponding to a plurality of time nodes, but the vehicle operation data should be as real-time as possible, that is, the vehicle operation parameters of the current time node or the last time node are included. The vehicle operation data is input into the thermal runaway prediction model which is trained and meets the requirements to perform real-time thermal runaway prediction on the battery.

[0106] In some embodiments, step S6 comprises: According to the determined set time window size n', the vehicle operation parameters generated by the vehicle at the current time node and the vehicle operation parameters corresponding to the previous n'-1 historical time nodes are divided into the same time window, and the vehicle operation parameters in the time window are input into the thermal runaway prediction model to perform real-time thermal runaway prediction on the vehicle.

[0107] For example, when the finally determined set time window size n' is 2000, the vehicle operation parameters generated at the current time node and the historical 1999 time node data are input into the thermal runaway prediction model trained in step S4 after data standardization processing, and real-time prediction is performed; in a specific embodiment, the predicted results are cached, and the predicted results of the connected time nodes are compared. If all show that the thermal runaway is abnormal, the vehicle can be warned of the risk of thermal runaway.

[0108] In this specification, the illustrative description of the above terms does not necessarily refer to the same embodiment or example. Moreover, the described specific features, structures, materials or characteristics can be combined in any one or more embodiments or examples in a suitable manner.

[0109] Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A method for predicting vehicle thermal runaway based on time series data, characterized in that: include: Acquiring training data, wherein the training data is time series data of vehicle operating parameters; Perform data slicing on the training data, wherein the training data is sliced ​​according to the time series and the set time window size; Establish a neural network-based thermal runaway prediction model, wherein the input of the thermal runaway prediction model is the vehicle operating parameters in a single slice data, and the output is a thermal runaway probability value; the thermal runaway prediction model includes a plurality of sequentially connected input layers, wherein the vehicle operating parameters of the last time node in a set time window are input to the first input layer, and the vehicle operating parameters of the remaining time nodes in the set time window are sequentially input into the remaining input layers in a time series as additional inputs until all time node data in the set window are exhausted; Several slice data and their corresponding thermal runaway probability value labels are used as sample data, and the thermal runaway prediction model is trained using the sample data; Acquiring vehicle operation data, where the vehicle operation data and the training data are both used to characterize the timing states of the vehicle and the battery pack during operation; The vehicle operation data is input into the trained thermal runaway prediction model to perform real-time thermal runaway prediction on the battery.

2. The method for predicting vehicle thermal runaway based on time series data according to claim 1, characterized in that: The time window is set to include n time nodes, and the thermal runaway prediction model includes n-1 input layers; The main input of the first input layer includes the vehicle operating parameters at the nth time node, and the vehicle operating parameters at the n-1th time node serve as the additional input of the first input layer; Each input layer from the 2nd to the n-1th input layer is defined as an intermediate input layer, then the main input of the i-th intermediate input layer includes the output of the i-1th input layer, and the vehicle operating parameters at the ni-th time node are used as the additional input of the i-th intermediate input layer; where i is greater than or equal to 2 and less than or equal to n-1.

3. The method for predicting vehicle thermal runaway based on time series data according to claim 2, characterized in that: Each input layer of the thermal runaway prediction model includes m neurons. The number of neurons in each input layer is equal to the number of characteristic data types in the vehicle operating parameters. In the first input layer, the total input of each neuron is: in, represents the total input of the jth neuron in the first input layer, ; Represents the j-th type of characteristic data in the vehicle operating parameters at the n-th time node; express Parameters; Represents the j-th characteristic data of the vehicle operating parameters at the n-1th time node; express Parameters.

4. The method for predicting vehicle thermal runaway based on time series data according to claim 3, characterized in that: In the middle input layer, the total input to each neuron is: in, represents the total input of the jth neuron in the i-th input layer; represents the output of the jth neuron in the i-1th input layer; express Parameters; Represents the j-th type of characteristic data in the vehicle operating parameters at the ni-th time node; express Parameters.

5. The method for predicting vehicle thermal runaway based on time series data according to claim 4, characterized in that: Each neuron in the input layer uses the sigmoid activation function to generate output. The expression of the sigmoid activation function is: in, represents the output of the neuron; y represents the total input of the neuron; and e represents the natural exponent.

6. The method for predicting vehicle thermal runaway based on time series data according to claim 4, characterized in that: The thermal runaway prediction model further includes several hidden layers, which are sequentially connected to the n-1th input layer. The total input amount of neurons in the hidden layer connected to the n-1th input layer, i.e., the first hidden layer, is expressed as: in, represents the total input of the jth neuron in the hidden layer connected to the n-1th input layer; represents the output of the jth neuron in the n-1th input layer; express Parameters; The total input to the neurons in the remaining hidden layers is expressed as: in, represents the total input of the jth neuron in the x+1th hidden layer, , X represents the total number of hidden layers; represents the output of the jth neural unit in the xth hidden layer; express Parameters; Each neuron in the hidden layer uses a sigmoid activation function to generate an output.

7. The method for predicting vehicle thermal runaway based on time series data according to claim 1, characterized in that: The vehicle operating parameters include: charging state, high voltage state, operating mode, vehicle speed, accumulated mileage, total voltage inside the battery pack, total current of the battery pack, insulation value of the battery pack, maximum cell voltage of the battery pack, minimum cell voltage of the battery pack, maximum temperature of the battery pack, minimum temperature of the battery pack, total number of batteries, total number of temperature probes, and part or all of the temperature value of each temperature probe; Before using the training data for model training and using the vehicle operation data for real-time thermal runaway prediction, the vehicle operating parameters are standardized. Among them, dummy variable conversion is used for data conversion of the charging state, high-voltage state and operating mode. The minimum-maximum value method is used for data conversion of the vehicle speed, cumulative mileage, total voltage inside the battery pack, total current of the battery pack and insulation value of the battery pack. The z-score method is used for data conversion of the maximum cell voltage of the battery pack, the minimum cell voltage of the battery pack, the maximum temperature of the battery pack, the minimum temperature of the battery pack, the total number of batteries, the total number of temperature probes and the temperature value of each temperature probe.

8. The method for predicting vehicle thermal runaway based on time series data according to claim 1, characterized in that: The loss function of the thermal runaway prediction model adopts a cross entropy loss function.

9. The method for predicting vehicle thermal runaway based on time series data according to claim 1, characterized in that: When using sample data to train the thermal runaway prediction model, the slice data generated by different set time window sizes are used to train the model respectively; Evaluate the prediction effect of the trained thermal runaway prediction model; evaluate the prediction effect of the model with different set time window sizes, and determine the set time window size selected for the final model based on the evaluation results; When using the thermal runaway prediction model to predict the real-time thermal runaway of the battery, the vehicle operating parameters generated by the vehicle at the current time node and the vehicle operating parameters corresponding to the previous n'-1 historical time nodes are divided into the same time window according to the determined set time window size n', and the vehicle operating parameters in the time window are input into the thermal runaway prediction model to perform real-time thermal runaway prediction on the battery.

10. The method for predicting vehicle thermal runaway based on time series data according to claim 9, characterized in that: The evaluation of the prediction effect of the trained thermal runaway prediction model includes: Use the test set to test the prediction effect of the trained model and calculate the accuracy, precision, and recall of the prediction. The calculation methods include: in, Indicates accuracy; Indicates the accuracy rate; represents the recall rate; Indicates the number of test results in which the actual thermal runaway label corresponding to the slice data is positive and the model prediction result is also positive; Indicates the number of test results in which the actual thermal runaway label corresponding to the slice data is negative and the model prediction result is also negative; Indicates the number of test results in which the actual thermal runaway label corresponding to the slice data is negative and the model prediction result is positive; Indicates the number of test results in which the actual thermal runaway label corresponding to the slice data is positive and the model prediction result is negative.

Images