LU500227B1 - A system and a method for smart grading and early warning the thermal runaway of a lithium-ion battery - Google Patents

Abstract

The present invention provides a system and a method for smart grading and early warning the thermal runaway of a lithium-ion battery. The method includes the following steps: use a fault-free battery characteristic element to train the thermal runaway prediction ability of a prediction model consisting of a long-term and short-term memory network, a time convolutional network and a GRU neural network; then input a battery characteristic element of the battery to be monitored in a time sequence to the prediction model, obtain a prediction value under a normal status at a certain moment, and then compare the prediction value with a collected real value corresponding to the moment. The larger a difference between the prediction value and the collected real value, the greater the risk of the thermal runaway of the battery. This operation not only combines respective advantages, but also avoids the problem of excessive coefficient deviation trained by respective models, so as to achieve high-precision thermal runaway of the battery grading detection, and improve the speed and accuracy of the thermal runaway warning of the lithium-ion battery.

Description

Description 7500227 A system and a method for smart grading and early warning the thermal runaway of a lithium-ion battery

TECHNICAL FIELD The present invention relates to the field of battery safety early warning technology, and in particular to a system and a method for smart grading and early warning the thermal runaway of a lithium-ion battery.

BACKGROUND TECHNOLOGY In recent years, an energy storage system of a lithium-ion battery has been widely used in various areas such as a new energy connected grid, a micro power grid, and a smart grid, and so on. The lithium-ion battery has broad prospects in large-scale energy storage systems because of high energy density, high conversion efficiency and fast response characteristics of the lithium-ion battery. However, with the large-scale application of the energy storage system of the lithium-ion battery, the safety of energy storage system has attracted people's attention. According to the statistical investigation of accidents, the thermal runaway of the lithium-ion battery because of the chemical reaction of the lithium-ion battery or an external influence has become a main cause of safety problems. After the thermal runaway of the lithium-ion battery, a fire has the characteristics of fast diffusion, high flame intensity, and a large amount of toxic gases. Even after the fire is extinguished, the fire may still reignite, which seriously threatens the entire energy storage system.

Because of the battery's own heat production and abnormal temperature rise, the thermal runaway of the battery causes a chain exothermic reaction, resulting the battery to burn and explode. According to the trigger conditions, the causes of the thermal runaway of the battery are mainly divided into three categories: mechanical abuse, such as mechanical deformation caused by acupuncture, squeezing, and impact of heavy objects; electrical abuse, such as lithium dendrites generated inside the battery to penetrate a diaphragm that are caused mainly by electrical component failures, such as overcharge and over-discharge ; and thermal abuse, such as decomposition of materials of positive and negative electrodes caused by excessive temperature. Therefore, the accurate prediction of lithium-ion thermal runaway failures will accelerate the development of energy storage technology, which is of great significance for promoting the transformation of China's new energy structure, ensuring energy security, and achieving energy conservation and emission reduction goals. Patent No. CN20160282373.1 disclosed a method and a device for detecting a battery failure. The method uses a data mining algorithm to identify a battery cell that fails in a battery pack, according to a battery characteristic element of the battery pack. The accuracy of battery failure detection can be improved without the need to use a preset threshold value based on manual experience. However, this method can only perform fault detection on a current battery status, cannot achieve early warning, and is prone to problems of untimely early warning and response.

1

In view of this, it is necessary to design an improved method and system for smart grad¥®80227 and early warning the thermal runaway of a lithium-ion battery to solve the forgoing problems.

SUMMARY OF INVENTION In order to overcome the forgoing shortcomings of the prior art, the objective of the present invention is to provide a system and a method for smart grading and early warning the thermal runaway of a lithium-ion battery. The method uses a long-term and short-term memory network, a time convolutional network and a GRU neural network to predict separately, and perform equal- weight coupling, which not only combines respective advantages, but also avoids the problem of excessive coefficient deviation trained by respective models, so as to achieve high-precision thermal runaway of the battery grading detection, and improve the speed and accuracy of the thermal runaway warning of the lithium-ion battery.

In order to achieve the forgoing objectives of the present invention, the present invention provides a method for smart grading and early warning the thermal runaway of a lithium-ion battery, including the following steps: S1. Collect a fault-free battery characteristic element, and train the prediction ability of a long- term and short-term memory network, a time convolutional network and a GRU neural network on a battery characteristic element in a time sequence, respectively.

S2. Collect the battery characteristic element of a battery to be monitored in the time sequence, input the battery characteristic element into the trained long-term and short-term memory network, the trained time convolutional network, and the trained GRU neural network, respectively, to obtain three sets of prediction values, and perform equal-weight coupling for the three sets of prediction values as fault-free prediction values.

S3. Continue to collect the battery characteristic element of the battery to be monitored in the time sequence as a real value, and compare the real value with the fault-free prediction value obtained in Step S2, wherein the greater a difference between the real value and the fault-free prediction value, the greater the risk of the thermal runaway of the battery.

As a further improvement of the present invention, the battery characteristic element includes but is not limited to one or more of temperature, current, voltage, and status of charge of the battery.

As a further improvement of the present invention, in Step S1, the fault-free battery characteristic element is obtained by filtering out the battery characteristic element of the thermal runway via a Gaussian filter.

As a further improvement of the present invention, the Gaussian filter is a low-pass filter with +200w of a frequency domain width and I1/100 of a threshold value.

As a further improvement of the present invention, in Step S1, training the prediction ability comprises: use the fault-free battery characteristic element in the time sequence to train and couple the long-term and short-term memory network, the time convolutional network, and the GRU neural network, respectively, to obtain the prediction result, that is, a change trend value of the battery characteristic element in time; and then return the respective loss values of the three networks to continue to train.

As a further improvement of the present invention, the loss values of the long-term and short- term memory network and the GRU neural network is obtained via an MSE loss function; the loss value of the time convolutional network is obtained by a mean absolute deviation MAD.

As a further improvement of the present invention, Step S3 further comprises: use a ratio of an absolute value of the difference between the real value and the fault-free prediction value to the 2 prediction value as an early warning coefficient of the thermal runaway, wherein the larger the 4900227 warning coefficient of the thermal runaway, the greater the risk of the thermal runaway of the battery.

As a further improvement of the present invention, Step S3 further comprises: divide the early warning coefficient of the thermal runaway into a plurality of grade intervals for grading the early warning; wherein the larger an upper limit of the grade interval, the greater the risk of the thermal runaway of the battery.

In order to achieve the forgoing objectives of the present invention, the present invention further provides a method and a system for smart grading and early warning the thermal runaway of a lithium- ion battery, including: a data collection module, configured to collect characteristic element data of the battery; a thermal runaway model building module, configured to construct a long-term and short-term memory network, a time convolutional network, and a GRU neural network as a prediction model associated with the characteristic element of the battery, and use the characteristic element of the battery to train the prediction ability of the characteristic element; a thermal runaway early warning module, configured to input the battery characteristic elements of a battery to be monitored in a time sequence into a prediction model to obtain a prediction value at a certain moment, and then compare the prediction value with the collected real value corresponding to the moment to obtain the early warning result of the thermal runaway.

As a further improvement of the present invention, the method and the system for smart grading and early warning the thermal runaway of the lithium-ion battery further comprises an equal-weight coupling module, the equal-weight coupling module is configured to perform equal-weight coupling for the prediction values of the long-term and short-term memory network, the time convolutional network, and the GRU neural network.

The present invention has the following beneficial effects:

1. A method for grading and early warning the thermal runaway of a lithium-ion battery provided by the present invention uses a long-term and short-term memory network, a time convolutional network and a GRU neural network to predict separately, and perform equal-weight coupling, which not only combines respective advantages, but also avoids the problem of excessive coefficient deviation trained by respective models, and further reduces training periods, so as to achieve high- precision thermal runaway of the battery grading detection, and improve the speed and accuracy of the thermal runaway warning of the lithium-ion battery, which helps promote the smart progress of new energy storage safety and accelerate the application of the fourth industrial revolution in the energy and energy storage industries.

2. The method for grading and early warning the thermal runaway of the lithium-ion battery provided by the present invention uses a historical battery characteristic element of a normal battery as a training set, and gradually corrects and reduces a difference between real values of the predicted characteristic element and the historical battery characteristic element. Because a model is trained and detected using normal data, prediction values obtained based on real-time monitoring data are all prediction values under normal conditions. Then compared with the real value at this moment, the larger a difference between the prediction value and the real value, the greater the risk of the thermal runaway of the battery. In this way, it is possible to predict the working status of the battery in advance in a time sequence, and realize the early warning of the thermal runaway in a real sense.

3. The method and the system for smart grading and early warning the thermal runaway of the lithium-ion battery provided by the present invention, during model training, loss generated the prediction value and the real value by the training set is returned to enhance a training effect and reduce a training workload of the entire model.

3

BRIEF DESCRIPTION OF THE DRAWINGS LUS00227 FIG. 1 is a flowchart of a model training method of a method and a system for smart grading and early warning the thermal runaway of a lithium-ion battery of the present invention.

FIG. 2 is a block diagram of a composition structure of a method for grading and early warning the thermal runaway of a lithium-ion battery of the present invention.

FIG. 3 is a structural diagram of a long-term and short-term memory network.

FIG. 4 is a structural diagram of a time convolutional network.

FIG. 5 is a structural diagram of a GRU neural network.

FIG. 6 is an image of a probability density function of a normal distribution in a frequency domain.

DESTAILED DESCRIPTION OF THE PRESENTLY PREFERRED EMBODIMENTS In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention is described in detail below in conjunction with specific embodiments.

Here, it should also be noted that, in order to avoid obscuring the present invention because of unnecessary details, only a structure and/or processing steps closely related to the solutions of the present invention are shown in the specific embodiments, and the other details that are not relevant to the present invention are omitted.

Moreover, it should also be noted that, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements not only include those elements, but also included other elements that are not explicitly listed, or also include elements inherent to this process, method, article or device.

During the long-term power supply operation of a lithium-ion battery, because of its life problems or external environmental problems, temperature abnormality may occur, and serious fire accidents may occur. The occurrence of the temperature abnormality is usually not a sudden change, but usually manifests as a changing trend. If the temperature abnormality cannot be early warned and dealt with in time, the temperature abnormality may gradually develop into an uncontrollable thermal runaway accident. Based on this, the present invention provides a method and a system for smart grading and early warning the thermal runaway of a lithium-ion battery. According to the historical monitoring parameters of the lithium-ion battery, the parameters under the normal status at a subsequent time are accurately predicted, and then the parameters are compared with the real-time monitored parameters. If the real-time monitored parameters are too far from the predicted parameters at the corresponding time, it means that the battery may have a risk of the thermal runaway, thereby achieving the early warning.

Please referring to FIGS. 1 and 2, a method and a system for smart grading and early warning the thermal runaway of a lithium-ion battery provided by the present invention includes the following steps: S1. Collect the characteristic elements (a historical battery characteristic element)of a fault-free battery in A time sequence, and train prediction ability of the battery characteristic element of a long- term and short-term memory network, a time convolutional network and a GRU neural network in a time sequence (that is, according to the historical battery characteristic element, are a difference between real values of the predicted characteristic element and the historical battery characteristic _— ———

element is gradually corrected and reduced. The prediction value and the real value arbitra#?/0227 compared in a training set should be a value at the same time point, so as to achieve the accurate prediction of the change trend). In this way, the historical data of the normal battery are used to train and test the forgoing network model, so that the historical data can predict the battery characteristic element of the normal status at the subsequent time in advance. When the prediction is made, because the model uses normal data to perform training and testing, prediction values obtained based on the real-time monitored data are all prediction values under the normal status.

S2. collect the battery characteristic element of a battery to be monitored in the time sequence, input the battery characteristic element into the trained long-term and short-term memory network, the trained time convolutional network, and the trained GRU neural network, respectively, to obtain three sets of prediction values, and perform equal-weight coupling for the three sets of prediction values as fault-free prediction values. In summary, equal-weight coupling can correct results, reduce a coefficient deviation trained by the respective model, and reduce the training period, so that the prediction results are more accurate.

S3. continue to collect the battery characteristic element of the battery to be monitored in the time sequence as a real value, and compare the real value with the fault-free prediction value obtained in Step S2, where the greater a difference between the real value and the fault-free prediction value, the greater the risk of the thermal runaway of the battery.

Step S3 further comprises: use a ratio of an absolute value of the difference between the real value and the fault-free prediction value to the prediction value as an early warning coefficient of the thermal runaway, where the larger the early warning coefficient of the thermal runaway, the greater the risk of the thermal runaway of the battery.

Step S3 further comprises: divide the early warning coefficient of the thermal runaway into a plurality of grade intervals for grading the early warning; where the larger an upper limit of the grade interval, the greater the risk of the thermal runaway of the battery. For example, according to a scale that an entire interval is 0-100% and each interval is 25%, the entire interval is divided into four grade intervals: [0-25%], [26-50%], [51-75%], and [76-100%]. When the interval value of the obtained thermal runaway warning coefficient is larger, the risk of thermal runaway is greater. Different degrees of warning are then carried out.

The battery characteristic element includes but is not limited to one or more of temperature, current, voltage, and status of charge of the battery. The temperature, the current, the voltage and the status of charge are preferable. Through concentrated training of a plurality of battery characteristic elements, the accuracy of predicting the normal status of the battery can be improved. With the method of the present invention, because an early warning model is trained based on the characteristic element of a non-thermal runaway fault battery, the model can sensitively give the prediction value of the battery characteristic element at a certain subsequent time based on the historical battery characteristic element of the normal battery. For example: if the battery works normally at time ti-ts, the battery has a tendency of thermal runaway at time ty, or is a starting point for the development of the thermal runaway. Then the prediction model consisting of the forgoing three networks can give the battery characteristic element in the normal status at the time t4 based on the battery characteristic element at the time t1-t3, compare the battery characteristic element with a monitored real value at the time ty to accurately determine abnormality at the time t4, and give an early warning immediately to take corresponding countermeasures. However, with the prior art, the starting point for the development of the thermal runaway at time t4 may not be able to sensitively determine that the battery is in a thermal runaway status, so that the early warning cannot be made.

In Step S1, the fault-free battery characteristic element is obtained by filtering out the battery characteristic element of the thermal runway via a Gaussian filter. That is, the Gaussian filter is 680227 to filter the collected singularity data (i.e., thermal runaway battery data) of the battery characteristic element. The Gaussian filter is a low-pass filter with £200w of a frequency domain width and I1/100 of a threshold value. For example, if a mean filter is used for approximation, the response function of the mean filter is as follows: x2 1 T2 g(x) = —=—e *° Please referred to FIG. 6, the response function is an image of a probability density function of a normal distribution in a frequency domain. 02 represents a variance of a normal distribution. As 6 increases, a graph is wider, and a peak value is lower. x is an independent variable, representing a main frequency (i.e., a peak value of the frequency) of the input time sequence that is Fourier- transformed into the frequency domain. To filter out a high-frequency thermal runaway battery signal, the Gaussian filter is used. As the main frequency increases, the weight g(x) is lower and close to zero, and the filtering can be completed.

In Step S1, training the prediction ability comprises: use the fault-free battery characteristic element in the time sequence to train and couple the long-term and short-term memory network, the time convolutional network, and the GRU neural network, respectively, to obtain the prediction result, that is, a change trend value of the battery characteristic element in time; and then return the respective loss values of the three networks to continue to train.

Specifically, please refer to FIGS 3 to 5. FIG. 3 is a structure diagram of the long-term and short- term memory network (LSTM). In FIG. 3, a structure in box A represents a cell status (ct), and a structure in box B represents a forgetting gate (ft), where 6 is a sigmoid function, and the output ft is at 0 ~1, representing the probability of forgetting. A structure in box C represents an input gate, where o is still the sigmoid function, and tanh is an activation function, which can be combined to update the cell status. A structure in box D represents cell renewal, where the input cell status is first multiplied by the output of the forget gate, and then added to a product of it and at of the input gate. A structure in box E represents a hidden status ht, which is the hidden status of the previous node, and the input signal of this node is multiplied by the cell status through the sigmoid function.

A loss function of LSTM is a MSE loss function: Kuna A Where, xi is a predicted element, and yi is an output prediction value of an actual element, which is returned as the square of the difference.

The use of the long-term and short-term memory network has the following advantages: (1) There is no occurrence of gradient disappearance or gradient explosion, and connection weight may be changed at each time step. (2) The forget gate and functions such as the sigmoid function and the tanh function participate in the internal self-circulation of nerve cells. For a system that processes a time sequence signal, the input of time information of an internal cell status of the LSTM is the output of a previous neuron, so it is not necessary that the outside manually decides what information the LSTM should forget or remember. The inside of the neuron is self-looping.

FIG. 4 is a structural diagram of a time convolutional network. A causal convolution TCN model is used. The TCN model includes one input layer, two hidden layers and one output layer, that is, the 6 value of the previous layer at the time t depends only on the value of the next layer at time {880227 before the time t. A first digit of each layer is filled with 0, is strictly one-way and cannot see future data. The first digit is a one-way structure, is not two-way, but is a strict time sequence model. The loss function uses an average absolute error MAD, as shown in the following formula: j 2, Co MAD = — 2 try mir).

Where, x; is the real value and m(x) is the prediction value.

The temporal convolutional network has the following characteristics: (1) The inherent characteristics of the time convolutional network enable TCN to extract different eigenvalues and eigenvectors in the time sequence signal.

(2) Parallelism of data processing. When the time sequence signal is given, the TCN can process the time sequence signal in parallel, without the need for sequential processing like RNN.

(3) The time convolutional network also rarely has the problems of gradient disappearance or gradient explosion (in the process of a traditional neural network processing the signal, because each layer has the influence of the activation function, it will cause cases that a plurality of characteristic weights continue to increase with an increase in the number of layers and a plurality of characteristic weights continue to decrease). Because a convolution kernel is mainly used for convolution, a fully connected layer uniformly maps different extracted characterstic vectors with weight.

(4) Lower memory usage. RNN needs to save information of each step when the RNN is used, which occupies a lot of memory. The convolution kernel of the TCN is shared in one layer, and the memory usage is lower.

FIG. 5 is a diagram of a GRU neural network structure. Compared with the LSTM, the GRU combines the forget gate and the input gate into a single update gate. The cell status and the hidden status are also mixed, and a plurality of other changes is added. A final model is simpler than a standard LSTM model. The loss function also uses the MSE function.

The GRU is a type of LSTM. The GRU combines the forget gate and the input gate into one update gate, which reduces many coefficients and calculation requirements. Although the accuracy of the GRU is not as good as that of the LSTM, the calculation is fast, and the GRU can avoid an error caused by the excessive calculation of the LSTM.

In this way, combining the forgoing three networks for equal weight coupling can not only correct the results, reduce the coefficient deviation of the training of the respective models, but also reduce a training period and make the prediction results more accurate.

The present invention further provides a method and a system for smart grading and early warning the thermal runaway of a lithium-ion battery, including: a data collection module, configured to collect characteristic element data of the battery; a thermal runaway model building module, configured to construct a long-term and short-term memory network, a time convolutional network, and a GRU neural network as a prediction model associated with the characteristic element of the battery, and use the characteristic element of the battery to train the prediction ability of the characteristic element; a thermal runaway early warning module, configured to input the battery characteristic elements of a battery to be monitored in a time sequence into a prediction model to obtain a prediction value at a certain moment, and then compare the prediction value with the collected real value corresponding to the moment to obtain the early warning result of the thermal runaway.

The method and the system for smart grading and early warning the thermal runaway of the _—

lithium-ion battery further comprises an equal-weight coupling module, the equal-weight coupl#R80227 module is configured to perform equal-weight coupling for the prediction values of the long-term and short-term memory network, the time convolutional network, and the GRU neural network. The system can not only correct the results, reduce the coefficient deviation of the training of the respective models, but also reduce a training period and make the prediction results more accurate.

In summary, the method and the system for grading and early warning the thermal runaway of the lithium-ion battery provided by the present invention uses the long-term and short-term memory network, the time convolutional network and the GRU neural network to predict separately, and perform equal-weight coupling, which not only combines respective advantages, but also avoids the problem of excessive coefficient deviation trained by respective models, and further reduces training periods, so as to achieve high-precision thermal runaway of the battery grading detection, improve the speed and accuracy of the thermal runaway warning of the lithium-ion battery, be able to predict the working status of the battery in advance in the time sequence, and realize the early warning of the thermal runaway in a true sense.

Finally, the foregoing embodiments are intended for describing instead of limiting the technical solutions of the present invention. Although the present invention is described in detail with reference to the foregoing preferred embodiments, the persons of ordinary skill in the art should understand that they may still make modifications or equivalent replacements to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

8

Claims (10)

Claims LU500227

1. À system and a method for smart grading and early warning the thermal runaway of a lithium- ion battery, characterizing by comprising the following steps: S1. collect a fault-free battery characteristic element, and train the prediction ability of a long- term and short-term memory network, a time convolutional network and a GRU neural network on a battery characteristic element in a time sequence, respectively; S2. collect the battery characteristic element of a battery to be monitored in the time sequence, input the battery characteristic element into the trained long-term and short-term memory network, the trained time convolutional network, and the trained GRU neural network, respectively, to obtain three sets of prediction values, and perform equal-weight coupling for the three sets of prediction values as fault-free prediction values; S3. continue to collect the battery characteristic element of the battery to be monitored in the time sequence as a real value, and compare the real value with the fault-free prediction value obtained in Step S2, wherein the greater a difference between the real value and the fault-free prediction value, the greater the risk of the thermal runaway of the battery.

2. The system and a method for smart grading and early warning the thermal runaway of a lithium-ion battery according to Claim 1, characterized in that the battery characteristic element includes but 1s not limited to one or more of temperature, current, voltage, and status of charge of the battery.

3. The system and a method for smart grading and early warning the thermal runaway of a lithium-ion battery according to Claim 1, characterized in that in Step S1, the fault-free battery characteristic element is obtained by filtering out the battery characteristic element of the thermal runway via a Gaussian filter .

4. The system and a method for smart grading and early warning the thermal runaway of a lithium-ion battery according to Claim 3, characterized in that the Gaussian filter is a low-pass filter with £200w of a frequency domain width and IT/100 of a threshold value.

5. The system and a method for smart grading and early warning the thermal runaway of a lithium-ion battery according to Claim 1, characterized in that in Step S1, training the prediction ability comprises: use the fault-free battery characteristic element in the time sequence to train and couple the long-term and short-term memory network, the time convolutional network, and the GRU neural network, respectively, to obtain the prediction result, that is, a change trend value of the battery characteristic element in time; and then return the respective loss values of the three networks to continue to train.

6. The system and a method for smart grading and early warning the thermal runaway of a lithium-ion battery according to Claim 5, characterized in that the loss values of the long-term and short-term memory network and the GRU neural network is obtained via an MSE loss function; the loss value of the time convolutional network is obtained by a mean absolute deviation MAD.

7. The system and a method for smart grading and early warning the thermal runaway of a lithium-ion battery according to Claim 1, characterized in that Step S3 further comprises: use a ratio of an absolute value of the difference between the real value and the fault-free prediction value to the prediction value as an early warning coefficient of the thermal runaway, wherein the larger the early warning coefficient of the thermal runaway, the greater the risk of the thermal runaway of the battery.

8. The A system and a method for smart grading and early warning the thermal runaway of a lithium-ion battery according to Claim 7, characterized in that Step S3 further comprises: divide the early warning coefficient of the thermal runaway into a plurality of grade intervals for grading the 9 early warning; wherein the larger an upper limit of the grade interval, the greater the risk of4R80227 thermal runaway of the battery.

9. A system and a method for smart grading and early warning the thermal runaway of a lithium- ion battery, characterizing by comprising: a data collection module, configured to collect characteristic element data of the battery; a thermal runaway model building module, configured to construct a long-term and short-term memory network, a time convolutional network, and a GRU neural network as a prediction model associated with the characteristic element of the battery, and use the characteristic element of the battery to train the prediction ability of the characteristic element; a thermal runaway early warning module, configured to input the battery characteristic elements of a battery to be monitored in a time sequence into a prediction model to obtain a prediction value at a certain moment, and then compare the prediction value with the collected real value corresponding to the moment to obtain the early warning result of the thermal runaway.

10. The system and a method for smart grading and early warning the thermal runaway of a lithium-ion battery according to Claim 9, characterized in that the system for grading and early warning the thermal runaway of the lithium-ion battery further comprises an equal-weight coupling module, the equal-weight coupling module is configured to perform equal-weight coupling for the prediction values of the long-term and short-term memory network, the time convolutional network, and the GRU neural network.