Thermal runaway warning method for energy storage batteries based on autoencoder and ensemble learning

Abstract

The present invention discloses a method for predicting thermal runaway of energy storage batteries based on autoencoders and ensemble learning. The steps are as follows: 1) Based on different training sample data, establish several basic models for predicting thermal runaway of energy storage batteries based on autoencoders; 2) Conduct ensemble training on several basic models for predicting thermal runaway of energy storage batteries based on autoencoders to obtain a model for predicting thermal runaway of energy storage batteries; 3) Obtain the real-time operation data of the energy storage batteries and input them into each basic model for predicting thermal runaway of energy storage batteries based on autoencoders respectively to obtain the judgment results of the thermal runaway state of the energy storage batteries output by each basic model for predicting thermal runaway of energy storage batteries; 4) Input all the judgment results of the thermal runaway state of the energy storage batteries into the model for predicting thermal runaway of energy storage batteries to calculate the probability of thermal runaway of the energy storage batteries. The present invention realizes early warning of potentially thermally runaway energy storage batteries by establishing an end-to-end data model, making up for the defects of experimental methods and mechanism model methods in this field.

Description

Technical Field

[0001] The present invention relates to the field of energy storage battery safety early warning, and specifically to a method for early warning thermal runaway of energy storage batteries based on autoencoders and ensemble learning. Background Art

[0002] The causes of thermal runaway problems in energy storage batteries are divided into three categories: mechanical failures, electrical failures, and thermal failures. Their common link is internal short circuit. The development time scale of internal short circuit reaches several hundred hours, and the phenomena in the initial stage are not obvious, while combustion and explosion will be triggered in a short time in the final stage. Therefore, the problem of thermal runaway prevention and control early warning is of great significance. To solve the problem of thermal runaway prevention and control early warning, the mainstream methods in the academic circle are divided into three categories: experimental-based methods, model-based methods, and data-driven methods. Experimental-based methods can obtain the battery temperature safety boundary and then guide the design of the battery system. However, the accuracy of experimental methods depends on a large number of experimental times, and there are problems such as potential safety hazards and high economic costs, making it difficult to popularize and apply. Model-based methods mainly indirectly warn of thermal runaway by estimating the temperature rise, voltage change, or temperature distribution of the battery. Its advantage is that the physical meaning is clear, but the existing model methods are targeted at relatively single working conditions and are difficult to apply to real working conditions.

[0003] Data-driven methods are one of the emerging research hotspots in this field. By using historical data, a data model of the relationship between parameters such as the voltage, temperature, and SOC of energy storage batteries and their thermal runaway can be established. Its advantage is that it can reflect the real working conditions of the battery and avoid the research on the complex electrochemical mechanism inside the energy storage battery. Currently, most data-driven methods are supervised methods. Their research idea is to obtain data of energy storage batteries with thermal runaway under specific working conditions through experimental methods, use it as a label to train a neural network model, and then judge the thermal runaway situation of the energy storage battery. However, the above-mentioned supervised methods have the following problems: Although the scale of energy storage battery operation data is large, the proportion of thermal runaway battery data is small, resulting in a small sample problem, which in turn leads to low model accuracy. Therefore, supervised methods have limitations in this problem. Unsupervised methods can learn from datasets without label annotation and are insensitive to unbalanced data. This method is suitable for the problem of discriminating abnormal energy storage batteries containing a large number of normal samples. However, the data differences between energy storage batteries in the middle and early stages of thermal runaway and normal batteries are weak, and it is difficult to distinguish thermal runaway energy storage batteries only by using simple distance calculation methods for unsupervised clustering (such as the Kmeans method). Summary of the Invention

[0004] The purpose of the present invention is to provide a method for early warning thermal runaway of energy storage batteries based on autoencoders and ensemble learning, including the following steps:

[0005] 1) Based on different training sample data, establish several basic models for early warning thermal runaway of energy storage batteries based on autoencoders;

[0006] The training sample data includes the operating state data of the energy storage battery;

[0007] The operating state data of the energy storage battery includes voltage, current, state of charge, temperature, and statistical variable M t ; The statistical variables include the variance of the single-cell voltage of the energy storage battery, the average voltage, the maximum voltage, the minimum voltage, the variance of all single-cell temperatures, the maximum temperature, and the minimum temperature.

[0008] The training sample data has been standardized;

[0009] The data x after standardization * is as follows:

[0010]

[0011] In the formula, x μ is the sample mean; x σ is the sample standard deviation, and x is the data before standardization.

[0012] The encoder and decoder of the basic model for thermal runaway warning of energy storage batteries based on the autoencoder are respectively as follows:

[0013] y = h(X in ) (8)

[0014]

[0015] In the formula, y is the model output; h(*) and f(*) represent the encoding and decoding functions; X in = [V t , I t , SOC t , T t , M t represents the input data at time t; V t , I t , SOC t , T t , M t are respectively the total voltage, total current, state of charge of the battery pack, temperature of the battery pack, and statistical variable at time t of the battery pack; represents the output data of the encoder at time t; respectively represent the reconstructed data of the total voltage, total current, state of charge of the battery pack, temperature of the battery pack, and statistical variable at time t of the battery pack;

[0016] The optimization objective of the basic model for thermal runaway warning of energy storage batteries based on the autoencoder is as follows:

[0017]

[0018] In the formula, m is the dimension of X in and dimension; x in,i,t is the value of the feature vector at the i-th dimension and the t-th moment in X in ; x out,i,t is the value of the feature vector at the i-th dimension and the t-th moment in. T is the total number of moments.

[0019] The judgment result y of the thermal runaway state of the energy storage battery output by the basic model for early warning of thermal runaway of the energy storage battery is as follows:

[0020]

[0021] In the formula, K is the set threshold; y = 1 indicates that the energy storage battery is thermally out of control; y = 0 indicates that the energy storage battery is operating normally.

[0022] 2) Integrate and train several basic models for early warning of thermal runaway of energy storage batteries based on autoencoders to obtain an early warning model for thermal runaway of energy storage batteries;

[0023] The early warning model for thermal runaway of the energy storage battery is as follows:

[0024]

[0025] In the formula, P is the probability that the energy storage battery is judged to be thermally out of control; y k is the judgment result of the k-th basic model on the energy storage battery; n is the total number of basic models.

[0026] 3) Obtain the real-time operation data of the energy storage battery and input it into each basic model for early warning of thermal runaway of the energy storage battery respectively to obtain the judgment result of the thermal runaway state of the energy storage battery output by each basic model for early warning of thermal runaway of the energy storage battery;

[0027] 4) Input all the judgment results of the thermal runaway state of the energy storage battery into the early warning model for thermal runaway of the energy storage battery to calculate the thermal runaway probability of the energy storage battery.

[0028] It should be noted that in the present invention, the reconstruction error of the time series data of the energy storage battery is calculated by the autoencoder to define the difference degree between batteries, and the reconstruction error basic model is constructed using the idea that the reconstruction error of normal data is small and the reconstruction error of abnormal data is large. Then, an early warning framework for thermal runaway of electrochemical energy storage batteries based on ensemble learning is proposed to quantify the thermal runaway probability of the energy storage battery and enhance the stability of the model, so that the basic model can accurately select the thermally runaway energy storage battery in different validation sets. Finally, through the real operation data of the electric vehicle energy storage battery, the effectiveness of the traditional Kmeans method and the method proposed in the present invention under various indexes is compared.

[0029] The technical effects of the present invention are beyond doubt. By establishing an end-to-end data model, the present invention realizes early warning of potentially thermally runaway energy storage batteries, making up for the defects of experimental methods and mechanism model methods in this field. The present invention proposes a data-driven early warning framework for energy storage battery thermal runaway, defines the degree of difference between batteries by using the reconstruction error of energy storage battery data, forms a basic model for discrimination, and selects the basic model as an autoencoder, which can effectively extract the time-varying characteristics of parameters such as voltage, temperature, and SOC of energy storage batteries during the thermal runaway process. The present invention further proposes an energy storage battery thermal runaway early warning technology based on ensemble learning, quantifies the probability of energy storage battery thermal runaway, and through comparison with the basic model, the numerical example shows that the present invention reduces the variance of the model while ensuring the model accuracy and improves the model stability. Description of the Drawings

[0030] Figure 1 It is a diagram of the basic model based on reconstruction error;

[0031] Figure 2 It is an early warning framework for thermal runaway of electrochemical energy storage batteries based on ensemble learning;

[0032] Figure 3 It is a flowchart of the method for early warning of thermal runaway of energy storage batteries based on autoencoder and ensemble learning. Detailed Embodiments

[0033] The present invention will be further described below in conjunction with embodiments, but it should not be understood that the above-mentioned subject scope of the present invention is limited to the following embodiments. Without departing from the above-mentioned technical idea of the present invention, various substitutions and changes made according to ordinary technical knowledge and customary means in the art shall be included within the protection scope of the present invention.

[0034] Embodiment 1:

[0035] See Figures 1 to 3 , the method for early warning of thermal runaway of energy storage batteries based on autoencoder and ensemble learning includes the following steps:

[0036] 1) Based on different training sample data, establish several basic models for early warning of thermal runaway of energy storage batteries based on autoencoders;

[0037] The training sample data includes the operating state data of energy storage batteries;

[0038] The operating state data of the energy storage battery includes voltage, current, state of charge, temperature, and statistical variable M t ; The statistical variables include the variance of the single-cell voltage of the energy storage battery, the voltage mean value, the voltage maximum value, the voltage minimum value, the variance of all single-cell temperatures, the temperature maximum value, and the temperature minimum value.

[0039] The training sample data has been standardized;

[0040] The data x after standardization * is as follows:

[0041]

[0042] In the formula, x μ is the sample mean; x σ is the sample standard deviation, and x is the data before standardization.

[0043] The encoder and decoder of the basic model for thermal runaway warning of energy storage batteries based on the autoencoder are respectively as follows:

[0044] y = h(X in ) (14)

[0045]

[0046] In the formula, y is the model output; h(*) and f(*) represent encoding and decoding functions; X in = [V t , I t , SOC t , T t , M t represents the input data at time t; V t , I t , SOC t , T t , M t are respectively the total voltage, total current, state of charge of the battery pack, battery pack temperature, and statistical variable at time t of the battery pack; represents the output data of the encoder at time t; respectively represent the reconstructed data of the total voltage, total current, state of charge of the battery pack, battery pack temperature, and statistical variable at time t of the battery pack;

[0047] The optimization objective of the basic model for thermal runaway warning of energy storage batteries based on the autoencoder is as follows:

[0048]

[0049] In the formula, m is the dimension of X in and ; x in,i,t is the value of the feature vector of the i-th dimension at time t in X in ; x out,i,t is the value of the feature vector of the i-th dimension at time t in. T is the total number of time instants.

[0050] The judgment result y of the thermal runaway state of the energy storage battery output by the basic model for thermal runaway warning of energy storage batteries is as follows:

[0051]

[0052] Wherein, K is a set threshold value; y = 1 indicates thermal runaway of the energy storage battery; y = 0 indicates normal operation of the energy storage battery.

[0053] 2) Integrate and train several basic models for thermal runaway warning of energy storage batteries based on autoencoders to obtain a thermal runaway warning model for energy storage batteries;

[0054] The thermal runaway warning model for energy storage batteries is as follows:

[0055]

[0056] Wherein, P is the probability that the energy storage battery is judged to be in thermal runaway; y k is the judgment result of the k-th basic model on this energy storage battery; n is the total number of basic models.

[0057] 3) Obtain the real-time operation data of the energy storage battery and input it into each basic model for thermal runaway warning of the energy storage battery respectively to obtain the judgment result of the thermal runaway state of the energy storage battery output by each basic model for thermal runaway warning of the energy storage battery;

[0058] 4) Input all the judgment results of the thermal runaway state of the energy storage battery into the thermal runaway warning model of the energy storage battery to calculate the thermal runaway probability of the energy storage battery.

[0059] Embodiment 2:

[0060] A method for thermal runaway warning of energy storage batteries based on autoencoders and ensemble learning includes the following steps:

[0061] 1. Basic model for thermal runaway warning of energy storage batteries based on autoencoders

[0062] The present invention uses the reconstruction error of energy storage battery data to measure the thermal runaway risk of energy storage batteries, and the idea is as follows: First, a basic reconstruction model is trained using the operation data of normal batteries; subsequently, the battery operation status is discriminated by measuring the reconstruction error of the data to be discriminated. In the field of anomaly detection, there is an abundance of normal data, so there is no small sample problem for the basic model based on normal data. For normal battery data, since the previously trained basic reconstruction model has learned the corresponding features, its reconstruction error is small. For thermal runaway battery data, since the previously trained basic model has not learned the thermal runaway battery data, and there are differences between thermal runaway data and normal data, its reconstruction error is large. Therefore, it can be used as a discrimination basis, and the specific process is as follows. The operation status of an energy storage battery can be described by variables such as its voltage (V), current (I), state of charge (SOC), and temperature (T). The above variables can be obtained through real-time monitoring by a battery management system (BMS). Since these variables have different dimensions and large differences in numerical values, directly using the original data to train the basic model will result in numerical problems, which is not conducive to the training of the basic reconstruction model. Therefore, data standardization processing should be carried out on the energy storage battery samples first. There are some extreme data in the energy storage battery samples that deviate from the sample mean. The above data are usually not faults but are caused by changes in the physical state of the battery. The z-score method uses the overall information of the sample and is less affected by extreme data. Therefore, the z-score standardization method is selected in the present invention. The z-score standardization method is a common data standardization method, which uses the sample mean and standard deviation for data preprocessing, as shown in Equation (4).

[0063]

[0064] In the formula: x μ is the sample mean; x σ is the sample standard deviation.

[0065] The standardized variables are used as the input feature variables of the basic model, as Figure 1 shown. The model input is denoted as X in,t = [V t , I t , SOC t , T t , M t , where X in,t represents the input data at time t, including the total voltage V t of the battery pack at time t, the total current I t , the state of charge SOC t of the battery pack, the temperature T t of the battery pack, and the statistical variable M t . The present invention uses statistical methods for Mt Build. A set of energy storage battery data contains the voltage and temperature data of 96 battery cells, and M t includes statistical variables such as the variance, mean, maximum value, minimum value of all cell voltages and the variance, maximum value, minimum value of all cell temperatures at time t in a set of battery data. The output is the reconstructed data of the input features, denoted as The variable dimensions of the input and output are equal.

[0066] The basic model of the present invention selects an autoencoder. To build an autoencoder, the following three tasks need to be completed: build an encoder, build a decoder, and set an optimization objective to measure the information lost due to compression. The parameters of the encoder and decoder can be optimized by minimizing the loss function. The Adam optimizer is selected in the present invention.

[0067] The encoder and decoder are constructed as follows:

[0068] y = h(X in ) (20)

[0069]

[0070] Let the total data duration be T. Then the cumulative reconstruction error is described by the mean square error function and used as the optimization objective, as shown in Equation (3).

[0071]

[0072] In the formula: X in represents the input data within the T time period, represents the output data within the T time period. m is the dimension of X in and . x in,i,t is the feature vector value of the i-th dimension at time t in X in , that is, variables such as V t , I t . x out,i,t is the feature vector value of the i-th dimension at time t in .

[0073] After the above basic model based on the autoencoder is trained, it can be used to judge the thermal runaway of the energy storage battery. As Figure 1 shown, after the energy storage battery to be judged passes through the basic model calculation to obtain the reconstruction error, a threshold needs to be set for the reconstruction error as the discrimination criterion for the battery thermal runaway. In the present invention, the corresponding reconstruction error set is obtained by training different normal battery samples, the mean value and standard deviation of this set are calculated, and the reconstruction error value exactly greater than two standard deviations of the mean value is selected as the threshold, denoted as K. The specific discrimination rule is shown in Equation (5):

[0074]

[0075] In the formula: y represents the judgment result of the basic model for the energy storage battery to be judged. If its reconstruction error is greater than or equal to the set threshold K, the energy storage battery is judged to be in thermal runaway and assigned a value of 1; if its reconstruction error is less than the set threshold K, the energy storage battery is judged to be normal and assigned a value of 0. In summary, the basic model can judge thermal runaway by measuring the reconstruction error of the battery data to be judged.

[0076] 2. Electrochemical energy storage battery thermal runaway early warning framework based on ensemble learning

[0077] The stability of the basic model for the thermal runaway early warning of energy storage batteries is defined as whether the model can accurately select the thermally runaway batteries in different test data sets. Since the parameters of the basic model are greatly affected by the input sample set, it is difficult to ensure the stability of the basic model. And the basic model will be used multiple times in the thermal runaway early warning of energy storage batteries. If its stability is insufficient, it will increase the probability of misjudgment and missed judgment in some warning processes. Ensemble learning judges thermal runaway by synthesizing the judgment results of multiple basic models, which can achieve the purpose of enhancing the model stability. Therefore, to ensure the stability of the basic model, the present invention further proposes a thermal runaway early warning framework based on the idea of ensemble learning, as Figure 2 shown. By synthesizing the discriminant results of multiple basic models trained by different normal energy storage battery sample sets, the thermal runaway probability P can be obtained, as shown in Equation (6):

[0078]

[0079] In the formula: P is the probability that the energy storage battery is judged to be in thermal runaway; y k is the judgment result of the k-th basic model for this energy storage battery, which can be calculated by formula (5); n is the total number of basic models. Figure 2 In the shown ensemble model, the architectures of all basic models are the same, and the training sample sets of each basic model are taken from different normal energy storage battery data.

[0080] The stability of the basic model mentioned in the present invention can be described by the probability variance calculated by the ensemble model. Then the variance of the ensemble model is as shown in Equation (25):

[0081]

[0082] According to Equation (25), it can be seen that the variance of the discriminant result of the ensemble model is negatively correlated with n. The larger n is, the smaller the variance is, and the higher the model stability is. In practical applications, due to limited sample data, the value of n in the present invention is determined according to the actual situation of the sample set.

[0083] 3. Energy storage battery thermal runaway early warning method based on autoencoder and ensemble learning

[0084] The present invention proposes a data-driven thermal runaway warning method for energy storage batteries. The overall flowchart is as Figure 3 shown and is specifically described as follows:

[0085] Step 1: Data preprocessing. First, to exclude the influence of abnormal data in the battery thermal runaway outbreak stage on model discrimination, the data in the outbreak stage is deleted. To eliminate the influence of the dimension of each feature quantity when calculating the reconstruction error, equation (4) is used to normalize the feature quantities with different dimensions.

[0086] Step 2: Feature selection. Using the total voltage V t , total current I t , state of charge SOC of the battery pack t , temperature T of the battery pack t and statistical variable M t and other time series variables to jointly form multi-dimensional time series features.

[0087] Step 3: Training of the basic reconstruction model. Based on the training sample set composed of normal batteries, train the basic reconstruction error model, and the basic model is an autoencoder.

[0088] Step 4: Training of the integrated model. By selecting different training samples, repeat step 3 multiple times to form multiple basic models, and then form an integrated model.

[0089] Step 5: Calculation of the sample to be judged. Input a sample of an energy storage battery to be judged, use the integrated model to calculate the reconstruction error of the sample, given the reconstruction error threshold, the sub-model outputs y(0 / 1), and finally use equation (6) to calculate the warning probability.

[0090] Example 3:

[0091] Verification experiment on the thermal runaway warning method for energy storage batteries based on autoencoder and ensemble learning, the content is as follows:

[0092] 1 Sample acquisition and data preprocessing

[0093] The present invention collects the actual energy storage battery data of a domestic company, which involves 48 groups of batteries in total, including parameters such as voltage, current, state of charge, and temperature of each group of batteries. The time span of all of them is half a year, and the sampling frequency is 10 s / group. Among them, 2 groups of energy storage batteries caused combustion and explosion due to thermal runaway. The above data is preprocessed as follows:

[0094] First, the thermal runaway outbreak stage is extremely short, and the parameters of the energy storage battery change significantly. To eliminate the influence of abnormal data in the thermal runaway outbreak stage of the battery on model discrimination, the data in the outbreak stage are deleted. Then, since there are too many working conditions involved in the non-charging state of the energy storage battery, the model may misjudge energy storage batteries in different working conditions as abnormal. However, the working conditions of the energy storage battery are relatively single during the charging process, and the data characteristics are relatively stable. Therefore, the charging part of the data is selected. Finally, to eliminate the influence of the dimension of each feature quantity when calculating the reconstruction error, the features with different dimensions are normalized according to Equation (1).

[0095] 2 Model Settings

[0096] The numerical example of the present invention will compare the following methods (M0 - M2), and the setting purposes are shown in Table 2. Among them, M1 and M2 are the methods proposed in the present invention.

[0097] M0: Unsupervised clustering method (Kmeans method).

[0098] M1: Thermal runaway judgment model of energy storage battery based on unsupervised learning Figure 2 ), and the autoencoder neural network is selected for the model. The number of hidden layers is 4, which are divided into an encoding layer and a decoding layer. The number of neurons in the first layer of the encoding layer is 30, and the number of neurons in the second layer is 15; the number of neurons in the first layer of the decoding layer is 15, and the number of neurons in the second layer is 30. The number of neurons in the input and output layers is the same. The initial learning rate is selected as 0.001, and the optimizer selects an improved training algorithm with an adaptive learning rate.

[0099] M2: On the basis of M1, an ensemble learning framework is further adopted.

[0100] Table 1 Comparison and Setting Purposes of Methods

[0101]

[0102] 3 Index Settings

[0103] To evaluate the performance of the algorithm proposed in the present invention, the following explanations are made: The thermal runaway early warning of the energy storage battery is essentially a binary classification problem, that is, all the batteries to be discriminated are divided into two categories: normal batteries and thermally runaway batteries. The confusion matrix is a basic tool for evaluating the credibility of a binary classifier. The confusion matrix shown in Table 1 shows all possible classification results of the classifier, where TP represents that both the true category and the model judgment category are normal, TN represents that the true category is abnormal and the model discriminates it as normal, FP represents that the true value is abnormal and the model discriminates it as normal, and FN represents that the true value is normal and the model discriminates it as abnormal.

[0104] Table 2 Confusion Matrix

[0105]

[0106] Based on the confusion matrix, multiple evaluation metrics for the binary classifier can be obtained: namely, accuracy, precision, recall, and comprehensive metrics such as (8)-(11).

[0107] Ac = (TP + TN) / (TP + TN + FP + FN) (8)

[0108] Pr = TP / (TP + FP) (9)

[0109] Re = TP / (TP + FN) (10)

[0110] F1 = 2 × Pr × Re / (Pr + Re) (11)

[0111] The industrial community has high requirements for the above metrics. For example, if the recall rate Re does not meet the requirements, it means that there are many misjudged thermal runaway batteries, which seriously endanger the lives and safety of the people. In addition to the above metrics, the warning probability for real thermal runaway vehicles can also characterize the effectiveness of the integrated model. The larger the values of the above metrics, the better the classification performance of the model. In addition, the present invention sets a warning sorting metric for thermal runaway batteries, and obtains a ranking from high to low by calculating the reconstruction error magnitudes of all batteries in the same test set, and records the serial numbers of two groups of thermal runaway batteries. The lower this value, the better the model performance.

[0112] 4 Comparison of the effects between the basic model of the present invention and the unsupervised clustering method based on distance calculation

[0113] In M1, 20 groups of normal energy storage battery samples are randomly selected from 48 groups of battery data as the training set, and the test set data is the remaining 28 groups of energy storage battery data (including two groups of thermal runaway batteries). The test set data value of M0 is the same as that set by M1. M0 directly performs clustering analysis on the test set, and the results are shown in Table 3. The M0 method cannot find thermal runaway vehicles in the set examples, and its classification accuracy, classification precision, and classification recall are only 78.6%, 0%, and 0% respectively; while the above metrics of the M1 method are all significantly higher than those of M0, and its classification accuracy, classification precision, and classification recall can reach 92.9%, 50%, and 100% respectively. And the warning sorting shows that M1 can effectively find thermal runaway vehicles. The above results indicate that the data differences between energy storage batteries in the mid-early stage of thermal runaway and normal batteries are weak, and it is difficult to distinguish thermal runaway energy storage batteries only by using a simple distance calculation method for unsupervised clustering. The existing unsupervised clustering methods cannot effectively warn of the thermal runaway of energy storage batteries, and verify the effectiveness of the reconstruction basic model M1 proposed by the present invention.

[0114] Table 3 Comparison of warning result metrics between M0 and M1

[0115]

[0116] 5 Comparison of the stability between the integrated model and the basic model of the present invention

[0117] The integrated model can improve the stability of the basic model. To illustrate the stable performance of the integrated model, the following example is set for illustration: M1 randomly selects 20 groups of normal energy storage battery samples from 48 groups of battery data as the training set, and the remaining 28 groups of energy storage battery samples (including 2 groups of thermal runaway battery data) as the test set. The training set and test set of M2 are the same as those of M1, except that M2 uses each group of normal battery data to train a basic model, including a total of 20 basic models. M1 and M2 are each tested 3 times. For each of the 3 tests, 20 different groups of normal batteries are randomly selected as the training set, and the remaining 28 groups of batteries (including two thermal runaway batteries) are used as the test set. In each test, the training set and test set of M2 are the same as those of M1. The test results are shown in Table 4. For M1, the evaluation indicators fluctuate greatly in the 3 tests. Among them, the evaluation accuracy in the first and third tests is relatively low, while the evaluation accuracy in the second test is relatively high. In contrast, the evaluation indicators of M2 are relatively high in the 3 tests, and except for a slight fluctuation in the warning sorting index of thermal runaway batteries, other indicators remain unchanged. The above results show that the integrated model for thermal runaway warning of energy storage batteries has higher stability than the basic model. In addition, although the total training and testing time of M2 is significantly longer than that of M1, compared with the time for early warning of battery thermal runaway, the extra time spent by M2 can be ignored.

[0118] Table 4 Comparison of warning result indicators between M1 and M2

[0119]

[0120] To solve the problem that it is difficult to effectively warn of the thermal runaway of energy storage batteries, the present invention proposes a data-driven method for warning of the thermal runaway of energy storage batteries. Through the real operation data of electric vehicle energy storage batteries, the effectiveness of the proposed method is verified: First, the present invention proposes a basic model for warning of the thermal runaway of energy storage batteries based on an autoencoder, uses the reconstruction error of energy storage battery data to define the degree of difference between batteries, forms a discrimination basic model, and illustrates the effect of the reconstruction basic model by comparing the basic model of the present invention with the Kmeans clustering method. Furthermore, a technology for warning of the thermal runaway of energy storage batteries based on ensemble learning is proposed, which quantifies the probability of thermal runaway of energy storage batteries. Through comparison with the basic model, the example shows that the variance of the model is reduced while ensuring the model accuracy, and the model stability is improved.

Claims

1. A thermal runaway warning method for energy storage batteries based on autoencoders and ensemble learning, characterized in that Including the following steps: 1) Based on different training sample data, establish several basic models for thermal runaway warning of energy storage batteries based on autoencoders; 2) Conduct integrated training on several basic models for thermal runaway warning of energy storage batteries based on autoencoders to obtain a thermal runaway warning model for energy storage batteries; 3) Obtain the real-time operation data of the energy storage battery and input it into each basic model for thermal runaway warning of the energy storage battery respectively to obtain the judgment results of the thermal runaway state of the energy storage battery output by each basic model for thermal runaway warning of the energy storage battery; 4) Input all the judgment results of the thermal runaway state of the energy storage battery into the thermal runaway warning model of the energy storage battery to calculate the thermal runaway probability of the energy storage battery; The encoder and decoder of the basic model for thermal runaway warning of the energy storage battery based on the autoencoder are respectively shown as follows: y = h(X in ) (1) Where y is the model output; h(*) and f(*) represent encoding and decoding functions; X in = [V t , I t , SOC t , T t , M t represents the input data at time t; V t , I t , SOC t , T t , M t are the total voltage, total current, state of charge of the battery pack, temperature of the battery pack, and statistical variable at time t of the battery pack, respectively; represents the output data of the encoder at time t; represent the reconstructed data of the total voltage, total current, state of charge of the battery pack, temperature of the battery pack, and statistical variable at time t of the battery pack, respectively; Optimization objectives of the basic model for thermal runaway warning of energy storage batteries based on autoencoders are as follows: Where m is X in and dimension; x in,i,t is the feature vector value of the i-th dimension at time t in X in ; x out,i,t is the feature vector value at the t-th moment of the i-th dimension in; T is the total number of moments.

2. The method for predicting thermal runaway of energy storage batteries based on autoencoders and ensemble learning according to claim 1, wherein: The training sample data includes the operation state data of the energy storage battery; The operating state data of the energy storage battery includes voltage, current, state of charge, temperature, and statistical variable M t ; The statistical variables include the variance of the single-cell voltage of the energy storage battery, the average voltage, the maximum voltage, the minimum voltage, the variance of all single-cell temperatures, the maximum temperature, and the minimum temperature.

3. The method for predicting thermal runaway of energy storage batteries based on autoencoders and ensemble learning according to claim 2, wherein: The training sample data has been standardized; The standardized data x * is as follows: where x μ is the sample mean; x σ is the sample standard deviation, and x is the data before standardization.

4. The method for predicting thermal runaway of energy storage batteries based on autoencoders and ensemble learning according to claim 1, characterized in that: The judgment result y of the thermal runaway state of the energy storage battery output by the basic model for thermal runaway warning of the energy storage battery is shown as follows: In the formula, K is a set threshold; y = 1 indicates that the energy storage battery has thermal runaway; y = 0 indicates that the energy storage battery is operating normally.

5. The method for predicting thermal runaway of energy storage batteries based on autoencoders and ensemble learning according to claim 1, wherein The thermal runaway warning model of the energy storage battery is shown as follows: where P is the probability that the energy storage battery is judged to be in thermal runaway; y k is the judgment result of the k-th basic model on the energy storage battery; n is the total number of basic models.

Images