CN114460481A - Energy storage battery thermal runaway early warning method based on Bi-LSTM and attention mechanism - Google Patents

Energy storage battery thermal runaway early warning method based on Bi-LSTM and attention mechanism Download PDF

Info

Publication number
CN114460481A
CN114460481A CN202210103432.XA CN202210103432A CN114460481A CN 114460481 A CN114460481 A CN 114460481A CN 202210103432 A CN202210103432 A CN 202210103432A CN 114460481 A CN114460481 A CN 114460481A
Authority
CN
China
Prior art keywords
energy storage
storage battery
lstm
thermal runaway
attention mechanism
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202210103432.XA
Other languages
Chinese (zh)
Inventor
杨燕
黎俊伟
王国胤
于洪
余娟
杨知方
王博石
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Chongqing University of Post and Telecommunications
Original Assignee
Chongqing University of Post and Telecommunications
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Chongqing University of Post and Telecommunications filed Critical Chongqing University of Post and Telecommunications
Priority to CN202210103432.XA priority Critical patent/CN114460481A/en
Publication of CN114460481A publication Critical patent/CN114460481A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01R—MEASURING ELECTRIC VARIABLES; MEASURING MAGNETIC VARIABLES
    • G01R31/00—Arrangements for testing electric properties; Arrangements for locating electric faults; Arrangements for electrical testing characterised by what is being tested not provided for elsewhere
    • G01R31/36—Arrangements for testing, measuring or monitoring the electrical condition of accumulators or electric batteries, e.g. capacity or state of charge [SoC]
    • G01R31/385—Arrangements for measuring battery or accumulator variables
    • G01R31/387—Determining ampere-hour charge capacity or SoC
    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01R—MEASURING ELECTRIC VARIABLES; MEASURING MAGNETIC VARIABLES
    • G01R31/00—Arrangements for testing electric properties; Arrangements for locating electric faults; Arrangements for electrical testing characterised by what is being tested not provided for elsewhere
    • G01R31/36—Arrangements for testing, measuring or monitoring the electrical condition of accumulators or electric batteries, e.g. capacity or state of charge [SoC]
    • G01R31/392—Determining battery ageing or deterioration, e.g. state of health

Abstract

The invention belongs to the field of safety early warning of energy storage batteries, and particularly relates to a thermal runaway early warning method of an energy storage battery based on a Bi-LSTM and attention mechanism, which comprises the steps of constructing a characteristic excavation model based on the Bi-LSTM and the attention mechanism, respectively training n basic models, calculating the reconstruction error of each basic model, forming a basic error set by the reconstruction errors of the n basic models, and calculating a reconstruction error threshold value according to the reconstruction error set; comparing the reconstruction errors of the n basic models with a reconstruction error threshold, if the reconstruction errors of the basic models are smaller than the reconstruction error threshold, judging that the battery is normal by the basic models, and otherwise, judging that the battery is out of control thermally; synthesizing n basic models to calculate the thermal runaway probability P of the battery; when the thermal runaway probability P exceeds 70 percent, performing thermal runaway warning; the invention realizes the accurate learning of the data characteristics of the energy storage battery and effectively extracts the time-varying characteristic of the state information data of the energy storage battery in the thermal runaway process.

Description

Energy storage battery thermal runaway early warning method based on Bi-LSTM and attention mechanism
Technical Field
The invention belongs to the field of safety early warning of energy storage batteries, and particularly relates to a thermal runaway early warning method of an energy storage battery based on Bi-LSTM and an attention mechanism.
Background
Currently, energy storage batteries are mainly used in electric vehicles and energy storage power stations. With the large-scale popularization of electric vehicles and energy storage power stations, burning and explosion accidents caused by thermal runaway of energy storage batteries are rare. Currently, the insufficient safety of an energy storage battery in an energy system brings great hidden dangers to national economy and life safety.
The causes of the thermal runaway problem of the energy storage battery are divided into three categories: mechanical faults, electrical faults, and thermal faults. The common link is internal short circuit. The development time scale of the internal short circuit reaches hundreds of hours, the phenomenon is not obvious in the initial stage, and the combustion and explosion can be caused in the final stage in a short time, so that the thermal runaway prevention and control early warning problem has important significance. In order to solve the problem of prevention, control and early warning of thermal runaway, the main methods in the academic world are divided into three categories: experimental based methods, model based methods, data driven based methods. The battery temperature safety boundary can be obtained by an experiment-based method so as to guide the design of a battery system. However, the accuracy of the experimental method depends on a large number of experimental times, so that the problems of potential safety hazards and high economic cost exist, and the method is difficult to popularize and apply. The model-based method indirectly warns thermal runaway mainly by estimating temperature rise, voltage change or temperature distribution of the battery. The method has the advantages of clear physical significance, but the existing model method is single in working condition and difficult to be suitable for real working condition.
The data-driven method is one of the new research hotspots in the field, and a data model of the thermal runaway relation between the voltage, the temperature, the SOC and other parameters of the energy storage battery can be established by using historical data. At present, most methods based on data driving are supervised methods, and the research idea is to obtain thermal runaway energy storage battery data under specific working conditions through an experimental method, train a neural network model by using the thermal runaway energy storage battery data as a label, and further judge the thermal runaway condition of the energy storage battery. However, the above supervision method has the following problems: although the operation data scale of the energy storage battery is large, the thermal runaway battery occupies a small proportion of data, and the problem of small samples exists, so that the model precision is not high, and the problem of the supervision method is limited. The unsupervised method can be used for learning the data set without label marks, is insensitive to unbalanced data, and is suitable for the problem of judging the abnormal conditions of the energy storage battery containing a large number of normal samples. However, the difference between the data of the energy storage battery in the early and middle stages of thermal runaway and the data of the normal battery is weak, and the thermal runaway energy storage battery is difficult to distinguish by carrying out unsupervised clustering (such as a Kmeans method) only by using a simple distance calculation method. In conclusion, the unsupervised method capable of effectively capturing the data difference between the thermal runaway battery of the energy storage battery and the normal battery has great application potential in the electrochemical energy storage battery thermal runaway early warning.
Disclosure of Invention
Aiming at the problem of the technical bottleneck of the existing energy storage battery thermal runaway accurate early warning, the invention provides an energy storage battery thermal runaway early warning method based on a Bi-LSTM and attention mechanism, which comprises the following steps:
acquiring the running state information of the energy storage battery, and standardizing the acquired running state information of the energy storage battery by adopting a z-score standardization method; taking n normal battery data sets as n training data;
constructing a feature mining model based on Bi-LSTM and an attention mechanism, respectively training through n training data to obtain n basic models, calculating a reconstruction error of each basic model, forming a basic error set by the reconstruction errors of the n basic models, and calculating a reconstruction error threshold according to the reconstruction error set;
comparing the reconstruction errors of the n basic models with a reconstruction error threshold, if the reconstruction errors of the basic models are smaller than the reconstruction error threshold, judging that the battery is normal by the basic models, and otherwise, judging that the battery is out of control thermally;
and (3) synthesizing n basic models to calculate the thermal runaway probability P of the battery, wherein the probability P is expressed as:
Figure RE-GDA0003555317320000021
wherein, ykJudging the energy storage battery for the kth basic model to obtain a battery thermal runaway result;
and when the thermal runaway probability P exceeds 70 percent, performing thermal runaway warning.
Further, normalization was performed using the z-score normalization method:
Figure RE-GDA0003555317320000022
wherein, x is a sample value to be normalized; x is the number of*The normalized sample value is obtained; x is the number ofμIs the sample mean; x is the number ofσIs the sample standard deviation.
Further, the Bi-LSTM and attention mechanism-based feature mining model comprises a Bi-LSTM model and an attention mechanism, the Bi-LSTM model comprises two layers of long and short term memory networks, namely a forward-propagation long and short term memory network and a backward-propagation long and short term memory network, and the propagation process of the forward-propagation long and short term memory network comprises the following steps:
ft＝σ(Wfht-1+Ufxt+bf)；
it＝σ(Wiht-1+Uixt+bi)；
Figure RE-GDA0003555317320000031
Figure RE-GDA0003555317320000032
ot＝σ(Woht-1+Uoxt+bo)；
ht＝ot⊙tanh(Ct)；
wherein x istInputting a hidden node of a current sequence; ctThe node state of the current sequence hidden layer is obtained; h is a total oftThe forward hidden layer output is at the time t; f. oft、it、otIs an intermediate variable, Wf、Wi、Wo、WcRespectively, is an intermediate variable ft、it、otAnd
Figure RE-GDA0003555317320000033
corresponding weight matrix, bf、bi、bo、bcIs distinguished by an intermediate variable ft、it、otAnd
Figure RE-GDA0003555317320000034
corresponding bias parameters; u shapef、Ui、Uc、UoInput weight matrixes corresponding to a forgetting gate, an input gate, a sequence hidden node state and an output gate respectively; σ () is an activation function; as indicates the multiplication of the corresponding elements in the matrix.
Further, the hidden layer output at the moment t of the Bi-LSTM model is expressed as:
h't＝atht+bthi+ct；
wherein, h'tRepresenting the hidden layer output of the Bi-LSTM model at the t moment; a ist、btThe reverse hidden layer output weight at time t, ht、hiRespectively outputting hidden layers at the t moments of the forward time sequence and the reverse time sequence; c. CtA bias term is represented.
Furthermore, in a feature mining model based on Bi-LSTM and attention mechanism, the feature classification input into the model at each moment is subjected to weight distribution by using the attention mechanism, the obtained hidden layer is used as input data, and a new hidden layer output h' is obtained by updating "tThe method comprises the following steps:
h”t＝Ath't；
wherein A istTo focus on the weight matrix of the force layer.
Further, attention is paid to the weight moment of the mechanical layerArray AtElement alpha of ith row and j columni,jIs output h 'of a Bi-LSTM model t moment hidden layer'tThe similarity between the ith element and the jth element is expressed as:
Figure RE-GDA0003555317320000041
wherein, h'i,t、h'j,tRespectively represent Bi-LSTM hidden layer output h 'at time t'tThe ith and jth elements of (1); l is the hidden layer output dimension.
Further, the goal of the feature mining model training process based on Bi-LSTM and attention mechanism is to minimize the reconstruction error of the input and output data, and the goal is expressed as:
Figure RE-GDA0003555317320000042
wherein the content of the first and second substances,
Figure RE-GDA0003555317320000043
is input XinAnd output
Figure RE-GDA0003555317320000044
The reconstruction error of (2); m is input XinAnd output
Figure RE-GDA0003555317320000045
N is the total duration; x is the number ofin,i,tIs input XinThe element value at the moment of the ith dimension t; x is the number ofout,i,tIs an output
Figure RE-GDA0003555317320000046
The value of the element at time t in the ith dimension.
Further, the process of comparing the reconstruction error of the sample to be detected with the reconstruction error threshold value and judging whether to perform early warning includes:
Figure RE-GDA0003555317320000047
wherein, X'inRepresenting a certain energy storage battery input to be judged, and is represented as Xin＝[V,SOC,I,T,M]V represents the battery voltage of the energy storage battery, SOC represents the state of charge of the energy storage battery, I represents the current of the energy storage battery, T represents the temperature of the energy storage battery, and M represents the voltage variance, the average value, the maximum value and the minimum value of all the monomers in a group of battery data, namely V, SOC, I and T;
Figure RE-GDA0003555317320000048
indicates the output of a certain energy storage battery to be judged as
Figure RE-GDA0003555317320000049
The reconstruction characteristics are according to the input characteristics V, SOC, I, T and M; y represents the judgment result of the basic model on the energy storage battery to be judged; k denotes a reconstruction error threshold.
Further, the calculation of the reconstruction error threshold K includes: and calculating the average value and the standard deviation of the reconstruction error set corresponding to the input data set, and selecting the reconstruction error at the position of two standard deviations greater than the average value as a reconstruction error threshold value K.
The method can realize accurate learning of data characteristics of the energy storage battery, and effectively extract time-varying characteristics of data such as voltage, temperature and SOC of the energy storage battery in the thermal runaway process; in addition, the invention realizes the differentiated weight learning of each parameter of the energy storage battery at different moments by improving the basic model through the attention mechanism, and finally forms an integrated model, thereby further reducing the missing judgment and the error judgment of the thermal runaway energy storage battery.
Drawings
FIG. 1 is a schematic diagram of a long and short term memory network used in the present invention;
FIG. 2 is a schematic diagram of data feature mining of an energy storage battery based on Bi-LSTM and attention mechanism according to the present invention;
FIG. 3 is an energy storage battery feature mining base model based on Bi-LSTM and attention mechanism according to the present invention;
FIG. 4 is a flow chart of an energy storage battery thermal runaway early warning method based on the Bi-LSTM and attention mechanism according to the invention;
fig. 5 shows the early warning probability values of the first 10 bits of the early warning probability in the embodiment of the present invention.
Detailed Description
The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be obtained by a person skilled in the art without making any creative effort based on the embodiments in the present invention, belong to the protection scope of the present invention.
The invention provides an energy storage battery thermal runaway early warning method based on Bi-LSTM and an attention mechanism, which comprises the following steps:
acquiring the running state information of the energy storage battery, and standardizing the acquired running state information of the energy storage battery by adopting a z-score standardization method; taking n normal battery data sets as n training data;
constructing a feature mining model based on Bi-LSTM and an attention mechanism, respectively training through n training data to obtain n basic models, calculating a reconstruction error of each basic model, forming a basic error set by the reconstruction errors of the n basic models, and calculating a reconstruction error threshold according to the reconstruction error set;
comparing the reconstruction errors of the n basic models with a reconstruction error threshold, if the reconstruction errors of the basic models are smaller than the reconstruction error threshold, judging that the battery is normal by the basic models, and otherwise, judging that the battery is out of control thermally;
and (3) synthesizing n basic models to calculate the thermal runaway probability P of the battery, wherein the probability P is expressed as:
Figure RE-GDA0003555317320000061
wherein, ykJudging the energy storage battery for the kth basic model to obtain a battery thermal runaway result;
and when the thermal runaway probability P exceeds 70 percent, performing thermal runaway warning.
Example 1
In this embodiment, a thermal runaway early warning method for an energy storage battery based on Bi-LSTM and attention mechanism is further specifically described.
The invention provides an energy storage battery data characteristic mining technology based on Bi-LSTM and an attention mechanism, and the energy storage battery data characteristic mining technology is used as a basic model for judging the thermal runaway of an energy storage battery. The operating state of the energy storage battery can be described by variables such as battery voltage (V), current (I), state of charge (SOC), and temperature (T), which can be monitored in real time by a Battery Management System (BMS). Because the variables have different dimensions and have larger difference in numerical value, in order to ensure the training precision of the basic reconstruction model, the data standardization processing should be firstly carried out on the energy storage battery sample. Since the energy storage battery may have some polar data deviating from the sample mean value under the influence of the change of the physical state of the battery, and the z-score standardization method utilizes the overall information of the sample and is less influenced by the polar data, the z-score standardization method is selected herein. The z-score normalization method uses sample mean and standard deviation for data pre-processing, namely:
Figure RE-GDA0003555317320000062
in the formula: x is a sample value to be normalized; x is the number of*The normalized sample value is obtained; x is the number ofμIs the sample mean; x is the number ofσIs the sample standard deviation. And inputting the feature vector by taking the normalized variable as a basic model. The thermal runaway process of the energy storage battery usually involves a time span of tens of hours, so that the basic model needs to have the capability of processing long-time span information in the process of data reconstruction. The Bi-directional Long-Short-Term Memory network (Bi-LSTM) can efficiently and effectively capture the Long-time span information, so the invention selects the Bi-LSTM as a basic model and is obtained by combining two layers of Long-Short-Term Memory networks (LSTM). The LSTM is shown in figure 1. The concrete model is as follows:
ft＝σ(Wfht-1+Ufxt+bf)；
it＝σ(Wiht-1+Uixt+bi)；
Figure RE-GDA0003555317320000071
Figure RE-GDA0003555317320000072
ot＝σ(Woht-1+Uoxt+bo)；
ht＝ot⊙tanh(Ct)；
wherein: x is the number oftInputting a hidden node of a current sequence; ctThe node state of the current sequence hidden layer is obtained; h istOutputting for the hidden node of the current sequence; f. oft，it，otIs an intermediate variable; w, b are weight parameters and bias parameters, respectively. The first five terms in the above formula are described by LSTM (. cndot.) and represent the input xtAnd output htAnd further the hidden layer output at time t of the Bi-LSTM can be expressed as:
ht＝LSTM(Vt,It,SOCt,Tt,Mt,ht-1)；
hi＝LSTM(Vt,It,SOCt,Tt,Mt,hi-1)；
h't＝atht+bthi+ct；
in the formula: LSTM (-) represents the operation process of the LSTM network; h istFor the output of the forward hidden layer at the time t, the input quantity V at the time tt、It、SOCt、Tt、MtAnd the previous time hidden layer output ht-1Is calculated by taking into account the pastThe influence of the hidden layer output on the current hidden layer output, the forward hidden layer is shown as a second layer circle from bottom to top in fig. 3; h isiFor reverse hidden layer output at t moment, the input quantity V at t momentt、It、SOCt、Tt、MtAnd the hidden layer output h at the next timei-1Calculated, taking into account the effect of the future hidden layer output on the current hidden layer output, the reverse hidden layer is shown as a circle in the third layer from bottom to top in fig. 3. Bi-LSTM utilizes reconstruction error as loss function and updates network hidden layer output h'tAs indicated by the fourth-level circle from bottom to top in fig. 3.
When the accumulated reconstruction errors in the N time period of the energy storage battery are calculated, the relative importance degrees of all input variables of the energy storage battery in the reconstruction errors at the current moment are different, the reconstruction errors of the energy storage battery multi-dimensional time sequence are directly considered to be inaccurate in accumulation of the reconstruction errors judged by the single-variable time sequence data abnormity, and the Bi-LSTM hidden layer at the current moment needs to be endowed with dynamic weight along with the time change. And the attention mechanism can realize weight difference learning of the characteristics of the time sequence data at different moments. Therefore, the attention mechanism is added into the Bi-LSTM basic reconstruction model, a color circle in the figure 3 represents an attention mechanism layer, and different weight coefficients are distributed to the feature vectors after the attention mechanism layer. Carrying out weight distribution on input features in the thermal runaway early warning basic model at each moment by using an attention mechanism, taking the obtained hidden layer as input data, and updating to obtain new hidden layer output h'tNamely:
h”t＝Ath't；
in the formula: a. thetTo focus on the weight matrix of the force layer. A. thetElement alpha in (1)i,jCalculated from the similarity, it is expressed as:
Figure RE-GDA0003555317320000081
in the formula: h'i,t，h'j,tRespectively representing Bi-LSTM hidden layer output h at t moment'tI and j elements of (1), l is the hidden layer output dimension. Finally, the output is obtained through the full connection layer
Figure RE-GDA0003555317320000082
Such as the uppermost circle in fig. 3.
In summary, in order to effectively extract data characteristics of the energy storage battery pack, such as voltage, temperature, current, SOC and the like, the invention comprehensively considers the time sequence characteristics of the energy storage battery operation data and the weight difference of the characteristics, so as to effectively mine the data characteristics of the energy storage battery. The model is used as a basic model for calculating the reconstruction error of the battery data, and the input vector of the model is recorded as Xin＝[V,SOC,I,T,M](ii) a The output is the reconstructed data of the input characteristics, and is recorded as
Figure RE-GDA0003555317320000083
The variable dimensions of the input and output are equal. The optimization goal of the model training process is to minimize the reconstruction error of the input and output data, which is expressed as:
Figure RE-GDA0003555317320000084
in the formula: j is input XinAnd output
Figure RE-GDA0003555317320000085
The reconstruction error of (2); m is XinAnd
Figure RE-GDA0003555317320000086
dimension; x is the number ofin,i,tIs XinThe value of the element at the moment t of the ith dimension; x is the number ofout,i,tIs composed of
Figure RE-GDA0003555317320000087
The value of the element at time t in the ith dimension. After the training of the basic model is completed, the method can be used for judging the thermal runaway of the energy storage battery to be detected. The energy storage battery to be detected obtains a reconstruction error after calculation of a basic model, and the judgment rule is as follows:
Figure RE-GDA0003555317320000091
in the formula: x'inRepresenting the input of a certain energy storage battery to be judged;
Figure RE-GDA0003555317320000092
representing the output of a certain energy storage battery to be judged; y represents the judgment result of the basic model on the energy storage battery to be judged; k represents a judgment threshold value for judging whether thermal runaway occurs; j' represents the reconstruction error of the sample to be judged calculated by the basic model. If the reconstruction error J' is larger than or equal to the set threshold value K, judging the energy storage battery to be thermal runaway, and assigning a value of 1; if the reconstruction error J' is smaller than the set threshold value K, the energy storage battery is judged to be normal, and the value is assigned to be 0. The method obtains a corresponding reconstruction error set by training different normal battery samples, calculates the average value and the standard deviation of the set, and selects the reconstruction error just larger than two standard deviations of the average value as a threshold value K.
In summary, the general flow chart of the energy storage battery thermal runaway early warning method based on the Bi-LSTM and attention mechanism of the present invention is shown in fig. 3, and specifically includes the following steps:
step 1: and (4) preprocessing data. Firstly, in order to eliminate the influence of the dimension of each feature vector when the reconstruction error is calculated, the different features of the dimension are normalized, and the z-score method is adopted in the invention. Then, a multi-dimensional time-series feature vector is constructed. And finally, taking n normal battery data sets in the obtained data as n training sets, and taking the rest battery data sets as a test set.
Step 2: and (5) training a model. Firstly, an energy storage battery data characteristic mining technology based on Bi-LSTM and attention mechanism in section 2 is selected to construct a basic model. Then, sequentially training n basic models based on the n training sets in the step 1; and obtaining a reconstruction error set, and calculating the reconstruction error set to obtain a reconstruction error threshold value K.
And step 3: and calculating a sample to be judged. Inputting a certain energy storage battery sample to be judged to the integrated model, and calculating the sample reconstructionComparing the error J' with a given reconstruction error threshold K, and judging the output y of each basic modelk(0/1), finally, calculating the early warning probability to obtain an early warning probability value P, namely:
Figure RE-GDA0003555317320000093
wherein, ykAnd judging that the energy storage battery is in thermal runaway for the kth basic model.
And regarding the energy storage battery with the early warning probability value exceeding 70% as a potential thermal runaway battery.
Example 2
In the embodiment, the actual energy storage battery data of a certain domestic company is collected, and the data totally relates to 60 groups of batteries, including parameters such as voltage, current, charge state and temperature of each group of batteries, wherein the time span is about half a year, and the sampling frequency is 10 Hz. Because the battery data in the non-charging state relate to various working conditions, and even among normal sample sets, the difference is large, so that the method only analyzes the actual charging data. However, since the charging frequency and the charging duration of each group of batteries are different, the shortest charging data length is used as a reference value, and the rest data lengths are used as reference values. There are 2 groups of energy storage batteries that have caused combustion and explosion due to thermal runaway. The data are preprocessed as follows:
firstly, the thermal runaway explosion stage is extremely short in time and unobvious in the phenomenon in the initial stage, and the combustion explosion is initiated in the final stage in a short time, each parameter of the energy storage battery is obviously changed in the process, and the explosion stage data is deleted in order to eliminate the influence of the abnormal data of the thermal runaway explosion stage of the battery on the model judgment;
then, due to the fact that the working conditions related to the non-charging state of the energy storage battery are too many, the model may misjudge the energy storage battery in different working conditions as abnormal. And the energy storage battery has single working condition in the charging process and relatively stable data characteristics, so that the charging part data of the energy storage battery is selected. Finally, in order to eliminate the influence of each characteristic dimension when the reconstruction error is calculated, normalization processing is carried out on different characteristics of the dimension.
The invention is compared to the following process (M1-M3) in which:
m1: the energy storage battery thermal runaway judgment model based on unsupervised learning selects a bidirectional long-time memory neural network, the number of hidden layers is 2, and the number of neurons in each layer is 40, so that the number of input and output neurons is consistent. The initial learning rate is selected to be 0.001, and the optimizer selects an improved algorithm of the self-adaptive learning rate.
M2: and selecting a fully-connected neural network as the model based on the unsupervised learning energy storage battery thermal runaway judgment model. Wherein, the hidden layer number is 3. The number of the first layer of neurons is 40, the number of the second layer of neurons is 15, and the number of the third layer of neurons is 40, namely, the number of the input and output layer neurons is ensured to be consistent. The initial learning rate is selected to be 0.0001, the optimizer selects a variable step gradient descent algorithm, and the basic models are integrated to form an integrated model.
M3: the model is a thermal runaway judgment model of the energy storage battery based on unsupervised learning, and a bidirectional long-time and short-time memory neural network is selected as the model. The number of hidden layers is 4. The first layer of the neuron number is 40, the second layer of the neuron number is 15, the third layer is different from M2, the fourth layer of the neuron number is 40 for the added attention mechanism layer, and the neuron number of the input and output layer is guaranteed to be consistent. The initial learning rate is selected to be 0.0001, the optimizer selects a variable step gradient descent algorithm, and the basic models are integrated to form an integrated model.
In the embodiment, six aspects of classification accuracy, classification recall, F1 value, early warning probability, and questionable vehicle suspicious degree sequencing are considered to measure the accuracy of the model. Wherein: the classification accuracy is the ratio of the correctly classified thermal runaway lithium battery to the correctly classified healthy lithium battery to the total number of batteries, namely:
Figure RE-GDA0003555317320000111
wherein A is the accuracy; trrightIs the number of correctly classified vehicles belonging to a true thermal runaway；NTrrightIs the number of healthy batteries correctly classified; trwrongIs the number of misclassified vehicles belonging to a true thermal runaway; NTrwrongIs the number of healthy batteries that are misclassified; the classification accuracy rate refers to the ratio of the number of the thermal runaway lithium batteries which are correctly classified to the number of the thermal runaway vehicles judged by the model, namely:
Figure RE-GDA0003555317320000112
wherein, P is the precision; trrightIs the number of correctly classified vehicles belonging to a real thermal runaway; trwrongThe number of the batteries which are judged to be thermal runaway by the model and are not thermal runaway temporarily; the recall rate refers to the ratio of the thermal runaway lithium batteries which are correctly classified to the number of the real thermal runaway lithium batteries, namely:
Figure RE-GDA0003555317320000113
in the formula: p is the precision; trrightThe number of correctly classified vehicles belonging to a true thermal runaway; NTrwrongThe number of the cells determined as normal cells among the real heat loss control cells. The F1 value is a harmonic mean of the precision and recall that takes into account the precision and recall in combination, namely:
Figure RE-GDA0003555317320000114
the number of times that the basic model in the definition integrated model of the early warning probability is judged to be 1 accounts for the proportion of all the basic models, and the judgment of the integrated model on the thermal runaway of a certain lithium battery is shown. The rank of the probability value of the problem vehicle in the current test batch is represented by the sequence of the suspicious degree of the problem vehicle, and whether the reconstruction error value of the model for the suspicious vehicle is different from that of a normal vehicle or not can be measured. The larger the above indexes are, the better the model classification performance is. In addition, the invention sets early warning sequencing indexes aiming at the thermal runaway batteries, sequences from high to low are obtained by calculating the reconstruction error of all batteries in the same test set, and the sequence numbers of two groups of thermal runaway batteries are recorded, and the lower the value is, the better the model performance is.
The method for accurately mining the data characteristics of the energy storage battery aims to improve the precision of each basic model in an integrated model. In order to illustrate the effect of the method on improving the model precision, the following calculation examples are set: m1 randomly selects 20 normal energy storage battery samples as a training set from 60 sets of battery data, and the remaining 40 energy storage battery samples (containing 2 sets of thermal runaway battery data) as a test set. The training set and the test set of M2 and M3 were set as M1, and the results are shown in table 2, table 3, and fig. 5.
TABLE 2 Integrated model Performance comparison based on different learning strategies
Figure RE-GDA0003555317320000121
TABLE 3 early warning probability and ranking of thermal runaway energy storage cells
Figure RE-GDA0003555317320000122
In summary, in order to solve the problem that the thermal runaway of the energy storage battery is difficult to effectively warn, three schemes of M1-M3 are provided for verification: firstly, the invention provides an energy storage battery parameter time sequence characteristic extraction technology based on Bi-LSTM to form an abnormal judgment basic model, defines the difference degree among batteries by using the reconstruction error of energy storage battery data, effectively extracts the time-varying characteristics of the energy storage battery voltage, temperature, SOC and other data in the thermal runaway process, realizes the differentiated weight learning of each parameter of the energy storage battery at different moments by improving the basic model through attention mechanism on the basis, reduces the missing judgment and the erroneous judgment of an integrated model on a thermal runaway vehicle, further forms the integrated model, and quantifies the thermal runaway probability of the energy storage battery. In the model of the invention, the thermal runaway energy storage battery can be obviously different from a normal battery, and the method has the highest early warning precision on the thermal runaway battery.
Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that changes, modifications, substitutions and alterations can be made in these embodiments without departing from the principles and spirit of the invention, the scope of which is defined in the appended claims and their equivalents.

Claims (9)

1. The energy storage battery thermal runaway early warning method based on the Bi-LSTM and the attention mechanism is characterized by comprising the following steps of:
acquiring the running state information of the energy storage battery, and standardizing the acquired running state information of the energy storage battery by adopting a z-score standardization method; taking n normal battery data sets as n training data;
constructing a feature mining model based on Bi-LSTM and an attention mechanism, respectively training through n training data to obtain n basic models, calculating a reconstruction error of each basic model, forming a basic error set by the reconstruction errors of the n basic models, and calculating a reconstruction error threshold according to the reconstruction error set;
comparing the reconstruction errors of the n basic models with a reconstruction error threshold, if the reconstruction errors of the basic models are smaller than the reconstruction error threshold, judging that the battery is normal by the basic models, and otherwise, judging that the battery is out of control thermally;
and (3) synthesizing n basic models to calculate the thermal runaway probability P of the battery, wherein the probability P is expressed as:
Figure FDA0003493049330000011
wherein, ykJudging the energy storage battery for the kth basic model to obtain a battery thermal runaway result;
and when the thermal runaway probability P exceeds 70 percent, performing thermal runaway warning.
2. The thermal runaway early warning method for the energy storage battery based on the Bi-LSTM and attention mechanism as claimed in claim 1, wherein the normalization is performed by a z-score normalization method:
Figure FDA0003493049330000012
wherein, x is a sample value to be normalized; x is the number of*The normalized sample value is obtained; x is the number ofμIs the sample mean; x is the number ofσIs the sample standard deviation.
3. The Bi-LSTM and attention mechanism based energy storage battery thermal runaway early warning method as claimed in claim 1, wherein the Bi-LSTM and attention mechanism based feature mining model comprises a Bi-LSTM model and an attention mechanism, the Bi-LSTM model comprises two layers of long-short term memory networks, namely a forward-propagation long-short term memory network and a backward-propagation long-short term memory network, and the propagation process of the forward-propagation long-short term memory network comprises:
ft＝σ(Wfht-1+Ufxt+bf)；
it＝σ(Wiht-1+Uixt+bi)；
Figure FDA0003493049330000021
Figure FDA0003493049330000022
ot＝σ(Woht-1+Uoxt+bo)；
ht＝ot⊙tanh(Ct)；
wherein x istInputting a hidden node of a current sequence; ctThe node state of the current sequence hidden layer is obtained; h istThe forward hidden layer output is at the time t; f. oft、it、otIs an intermediate variable, Wf、Wi、Wo、WcRespectively, is an intermediate variable ft、it、otAnd
Figure FDA0003493049330000023
corresponding weight matrix, bf、bi、bo、bcIs distinguished by an intermediate variable ft、it、otAnd
Figure FDA0003493049330000024
corresponding bias parameters; u shapef、Ui、Uc、UoInput weight matrixes corresponding to a forgetting gate, an input gate, a sequence hidden node state and an output gate respectively; σ () is an activation function; an indicator indicates that the corresponding element in the matrix is multiplied.
4. The Bi-LSTM and attention mechanism-based energy storage battery thermal runaway early warning method as claimed in claim 3, wherein the output of the hidden layer at the t moment of the Bi-LSTM model is represented as:
h't＝atht+bthi+ct；
wherein, h'tRepresenting the hidden layer output of the Bi-LSTM model at the t moment; a ist、btThe reverse hidden layer output weight at time t, ht、hiRespectively outputting hidden layers at the t moments of the forward time sequence and the reverse time sequence; c. CtA bias term is represented.
5. The method for warning thermal runaway of an energy storage battery based on a Bi-LSTM and attention mechanism as claimed in claim 4, wherein in the Bi-LSTM and attention mechanism based feature mining model, the feature classification input to the model at each moment is subjected to weight distribution by using the attention mechanism, the obtained hidden layer is used as input data, and a new hidden layer output h' is obtained by updating "tThe method comprises the following steps:
h”t＝Ath't；
wherein A istTo focus on the weight matrix of the force layer.
6. The thermal runaway early warning method for the energy storage battery based on the Bi-LSTM and the attention mechanism as claimed in claim 1, wherein the weight matrix A of the attention mechanism layertElement alpha of ith row and j columni,jIs output h 'of a Bi-LSTM model t moment hidden layer'tThe similarity between the ith element and the jth element is expressed as:
Figure FDA0003493049330000031
wherein, h'i,t、h'j,tRespectively represent Bi-LSTM hidden layer output h 'at time t'tThe ith and jth elements of (1); l is the hidden layer output dimension.
7. The thermal runaway early warning method for the energy storage battery based on the Bi-LSTM and the attention mechanism as claimed in claim 1, wherein the goal of the training process of the feature mining model based on the Bi-LSTM and the attention mechanism is to minimize the reconstruction error of the input and output data, and the goal is expressed as:
Figure FDA0003493049330000032
wherein the content of the first and second substances,
Figure FDA0003493049330000033
is input XinAnd output
Figure FDA0003493049330000034
The reconstruction error of (2); m is input XinAnd output
Figure FDA0003493049330000035
N is the total duration; x is the number ofin,i,tIs input XinThe element value at the moment of the ith dimension t; x is the number ofout,i,tIs an output
Figure FDA0003493049330000036
The value of the element at time t in the ith dimension.
8. The energy storage battery thermal runaway early warning method based on the Bi-LSTM and attention mechanism as claimed in claim 7, wherein the process of comparing the reconstruction error of the sample to be detected with the reconstruction error threshold and judging whether to perform early warning comprises:
Figure FDA0003493049330000037
wherein, X'inRepresenting a certain energy storage battery input to be judged, and is represented as Xin＝[V,SOC,I,T,M]V represents the battery voltage of the energy storage battery, SOC represents the state of charge of the energy storage battery, I represents the current of the energy storage battery, T represents the temperature of the energy storage battery, and M represents the voltage variance, the average value, the maximum value and the minimum value of all the monomers in a group of battery data, namely V, SOC, I and T;
Figure FDA0003493049330000038
indicates the output of a certain energy storage battery to be judged as
Figure FDA0003493049330000039
Figure FDA00034930493300000310
The reconstruction characteristics are according to the input characteristics V, SOC, I, T and M; y represents the judgment result of the basic model on the energy storage battery to be judged; k denotes a reconstruction error threshold.
9. The Bi-LSTM and attention mechanism based energy storage battery thermal runaway early warning method of claim 1, wherein the calculation of the reconstruction error threshold K comprises: and calculating the average value and the standard deviation of the reconstruction error set corresponding to the input data set, and selecting the reconstruction error at the position of two standard deviations greater than the average value as a reconstruction error threshold value K.
CN202210103432.XA 2022-01-27 2022-01-27 Energy storage battery thermal runaway early warning method based on Bi-LSTM and attention mechanism Pending CN114460481A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202210103432.XA CN114460481A (en) 2022-01-27 2022-01-27 Energy storage battery thermal runaway early warning method based on Bi-LSTM and attention mechanism

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202210103432.XA CN114460481A (en) 2022-01-27 2022-01-27 Energy storage battery thermal runaway early warning method based on Bi-LSTM and attention mechanism

Publications (1)

Publication Number Publication Date
CN114460481A true CN114460481A (en) 2022-05-10

Family

ID=81411786

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202210103432.XA Pending CN114460481A (en) 2022-01-27 2022-01-27 Energy storage battery thermal runaway early warning method based on Bi-LSTM and attention mechanism

Country Status (1)

Country Link
CN (1) CN114460481A (en)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114798503A (en) * 2022-06-07 2022-07-29 蜂巢能源科技股份有限公司 Lithium ion battery screening method and device based on safety and electronic equipment
CN116565402A (en) * 2023-07-11 2023-08-08 南京宁煜致科信息技术有限公司 Energy storage battery cooling system and control method

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114798503A (en) * 2022-06-07 2022-07-29 蜂巢能源科技股份有限公司 Lithium ion battery screening method and device based on safety and electronic equipment
CN114798503B (en) * 2022-06-07 2023-12-08 蜂巢能源科技股份有限公司 Safety-based lithium ion battery screening method and device and electronic equipment
CN116565402A (en) * 2023-07-11 2023-08-08 南京宁煜致科信息技术有限公司 Energy storage battery cooling system and control method
CN116565402B (en) * 2023-07-11 2023-09-29 南京宁煜致科信息技术有限公司 Energy storage battery cooling system and control method

Similar Documents

Publication Publication Date Title
CN112092675B (en) Battery thermal runaway early warning method, system and server
Zhang et al. Research on electric vehicle charging safety warning model based on back propagation neural network optimized by improved gray wolf algorithm
CN112379269B (en) Battery abnormality detection model training and detection method and device thereof
CN109375116B (en) Battery system abnormal battery identification method based on self-encoder
CN113158947B (en) Power battery health scoring method, system and storage medium
CN114460481A (en) Energy storage battery thermal runaway early warning method based on Bi-LSTM and attention mechanism
CN109604186A (en) Power battery performance flexibility assesses method for separating
CN113702843B (en) Lithium battery parameter identification and SOC estimation method based on suburb optimization algorithm
CN115366683A (en) Fault diagnosis strategy for new energy automobile power battery multi-dimensional model fusion
CN114559819A (en) Electric vehicle battery safety early warning method based on signal processing
CN115508711B (en) Electric vehicle power lithium battery residual life prediction method based on deep network model
CN114545274A (en) Lithium battery residual life prediction method
CN107895212A (en) Lead-acid battery life-span prediction method based on sliding window and various visual angles Fusion Features
CN111273125A (en) RST-CNN-based power cable channel fault diagnosis method
CN114565008A (en) Lithium battery abnormity detection method based on long-time and short-time memory self-encoder
CN115308628A (en) Battery health state monitoring method
CN113687242A (en) Lithium ion battery SOH estimation method for optimizing and improving GRU neural network based on GA algorithm
CN115659236A (en) Direct current charging pile fault early warning method based on soft classification
Yao et al. Fault identification of lithium-ion battery pack for electric vehicle based on ga optimized ELM neural network
Li et al. Timely thermal runaway prognosis for battery systems in real-world electric vehicles based on temperature abnormality
CN115469227A (en) Set variational self-encoder and dynamic regular lithium battery abnormity detection method
CN114036647A (en) Power battery safety risk assessment method based on real vehicle data
CN114580496A (en) Self-detection method for single fault of lithium ion battery
CN116609671A (en) Battery capacity prediction method, device, equipment and storage medium
CN114418465B (en) Data-driven power battery use behavior quantitative evaluation method

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination