CN116068399A - Lithium battery health state estimation method based on feature selection and time sequence attention - Google Patents
Lithium battery health state estimation method based on feature selection and time sequence attention Download PDFInfo
- Publication number
- CN116068399A CN116068399A CN202310082075.8A CN202310082075A CN116068399A CN 116068399 A CN116068399 A CN 116068399A CN 202310082075 A CN202310082075 A CN 202310082075A CN 116068399 A CN116068399 A CN 116068399A
- Authority
- CN
- China
- Prior art keywords
- charge
- discharge
- lithium battery
- value
- representing
- 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
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]
- G01R31/3644—Constructional arrangements
- G01R31/3648—Constructional arrangements comprising digital calculation means, e.g. for performing an algorithm
-
- 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]
- G01R31/367—Software therefor, e.g. for battery testing using modelling or look-up tables
-
- 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
-
- 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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Abstract
The invention discloses a lithium battery health state estimation method based on feature selection and time sequence attention, which comprises the following steps: 1, acquiring data sets of battery capacity, voltage, temperature and current of different rounds in multiple charge and discharge cycles of a lithium battery; 2, preprocessing charge-discharge related data to obtain health factor data with different characteristics; 2, carrying out pearson correlation coefficient analysis on different health factor data to obtain health factor data with higher correlation degree; 3, constructing a convolution time sequence attention network model M; and 4, sending the related data into a convolution time sequence attention network model M for model training, and finally obtaining a trained convolution time sequence attention network model M'. The method can effectively improve the accuracy of lithium battery health state prediction, is simple in model realization, easy in acquisition of training data, high in practicability, high in prediction accuracy and the like, and is suitable for online lithium battery health state prediction and lithium battery capacity estimation.
Description
Technical Field
The invention relates to a attention mechanism, a convolution and long-term memory network and lithium battery performance evaluation prediction, belongs to the field of regression prediction, and particularly relates to a lithium battery health state estimation method based on feature selection and time sequence attention.
Background
With the development of the prior battery technology and the importance of energy sources, the construction and research of the energy storage device are increasing at home and abroad. At present, lithium batteries and lead storage batteries are mainly adopted for electrochemical energy storage. The lithium iron phosphate battery has the advantages of high working voltage, high energy density, long cycle life, good safety performance, small self-discharge rate, no memory effect and the like, and is widely applied to the scenes of 5G base stations, electric automobiles, energy storage power stations and the like.
However, lithium batteries are dynamic chemical reaction discharge changes, and the lithium batteries can age under the influence of factors such as the use times, the ambient temperature, the charge-discharge current and voltage and the like, so that the internal resistance of the batteries is increased, and the temperature is abnormal and the like. Thereby affecting the service life. If the state of the battery cannot be predicted timely and accurately, once the energy storage battery pack has a problem, huge potential safety hazards and property loss can be brought. The State of health (SOH) indicates a State of decrease of the current battery capacity compared with the new battery capacity, and the capacity ratio of the new battery to the old battery is used as a criterion for determining whether the battery is eliminated. The research on the SOH of the lithium battery not only can quantify whether the current battery accords with the use standard, but also can ensure the stable operation of the energy storage system, and reduces the energy waste and the potential safety hazard. Therefore, if the health state of the lithium battery can be estimated timely and accurately, the method has important significance for management of the lithium battery and protection of the energy storage device.
The capacity and internal resistance of lithium batteries are commonly referred to as direct health factors, and are commonly used in SOH prediction of batteries, but measuring the capacity and internal resistance of lithium batteries requires high hardware conditions and the measurement process is time-consuming, which makes it extremely difficult to measure these data in real time. Therefore, if new health factors can be extracted from parameters such as voltage, current, temperature and the like which are easy to measure in real time to replace the capacity or the internal resistance of the lithium battery, the SOH prediction of the lithium battery is simpler and the SOH prediction result is more accurate.
At present, research on online state estimation of SOH of lithium batteries at home and abroad can be divided into two types based on model and data driving:
(1) Model-based SOH predictions are largely dependent on the impact on cell internal structure and material properties, and are implemented in combination with electrochemical decay mechanisms. The method expresses the internal structure of the battery and the physical and chemical reaction which occurs by using externally monitored parameters, constructs a physical degradation model according to the corresponding relation, and realizes SOH prediction of the lithium battery through parameter estimation of the model. The method mainly comprises the following steps: a degradation mechanism model; an equivalent circuit model; an empirical degradation model;
(2) The data-driven lithium battery SOH prediction method is the most widely used prediction method at home and abroad at present, does not need to consider complex chemical reactions occurring in the battery, does not need professional related electrochemical knowledge, and has higher universality. The SOH of the lithium battery is predicted based on a data driving method, and parameters which can be used for representing degradation characteristics of the lithium battery are mainly searched from charge and discharge data of the lithium battery. And based on the parameters, a deep learning or machine learning model is constructed, and then the prediction of the SOH of the lithium battery is realized through model training.
However, research now shows that there are a number of problems associated with the prediction of SOH in lithium batteries, such as: the model is not fine enough, all data characteristics cannot be represented, and good prediction precision cannot be obtained; training is performed by only using a single health factor, and the inherent relation of other data in the charge and discharge of the lithium battery cannot be learned; the model is too complex or too much data is used for training, the model is difficult to fit the data, and the model cannot be well deployed to practical application in real time.
Disclosure of Invention
The invention aims to solve the defects in the prior art, and provides a lithium battery health state estimation method based on feature selection and time sequence attention, so that the learning capacity and fitting performance of a model on data are improved by adding training data features, and the complexity of the model is reduced and the key features among learning data are emphasized, thereby obviously improving the accuracy of battery health state prediction of a lithium battery.
In order to achieve the aim of the invention, the invention adopts the following technical scheme:
the invention discloses a lithium battery health state estimation method based on feature selection and time sequence attention, which is characterized by comprising the following steps:
step one, acquiring data sets of battery capacity, voltage, temperature and current of different rounds in multiple charge and discharge cycles of a lithium battery, and preprocessing the data sets;
step 1.1, acquiring discharge voltage, charge current and discharge temperature data of the lithium battery under the z-th charge-discharge cycle, performing abnormal value removal treatment, and respectively performing uniform data length on the discharge voltage, charge current and discharge temperature data of the lithium battery after abnormal values are removedObtaining a charge-discharge dataset with uniform data length under a z-th charge-discharge cycle, comprising: discharge voltage U z ＝{u z1 ,u z2 ,u z3 …,u zj …,u zJ Charge current data I z ＝{i z1 ,i z2 ,…,i zj …,i zJ Charge-discharge temperature data T z ＝{t z1 ,t z2 ,t z3 …,t zj …,t zJ -a }; wherein u is zj Representing a jth discharge voltage value in a zth discharge cycle; i.e zj Representing a jth charging current value in a z-th charging cycle; t is t zj Representing a jth discharge temperature value in a zth discharge cycle; j represents the record times of one charge-discharge cycle in the charge-discharge data set;
step 1.2, respectively extracting maximum value and minimum value data of discharge voltage, charge current and discharge temperature data of the lithium battery in a z-th charge-discharge cycle, and carrying out averaging operation to obtain nine health factor data consisting of the maximum value, the minimum value and the average value of the charge-discharge data of the lithium battery in the z-th charge-discharge cycle, wherein the method comprises the following steps: voltage health factor U z ′＝{u z,max ,u z,min ,u z,avg -a }; current health factor I z ′＝{i z,max ,i z,min ,i z,avg -a }; temperature health factor T z ′＝{t z,max ,t z,min ,t z,avg }, where u z,max Representing the maximum value of the discharge voltage in the z-th discharge cycle; u (u) z,min Representing the minimum value of the discharge voltage in the z-th discharge cycle; u (u) z,avg Representing the average value of the discharge voltage in the z-th discharge cycle; i.e z,max Representing the maximum value of the charging current in the z-th charging cycle; i.e z,min Representing a minimum value of the charging current in the z-th charging cycle; i.e z,avg Representing an average value of the charging current in the z-th charging cycle; t is t z,max Representing the maximum value of the battery temperature in the z-th discharge cycle; t is t z,min Representing a minimum value of the battery temperature in the z-th discharge cycle; t is t z,avg Mean value of battery temperature in z-th discharge cycle;
step 1.3, willAfter the nine health factor data are spliced with the capacity of the lithium battery, normalization processing is carried out, so that preprocessed health factor data P= { P is obtained 1 ,p 2 ,…,p z …,p Z P is }, where z Health factor data representing the z-th charge-discharge cycle after pretreatment, and p z ＝{U z ′,I z ′,T z ′,C z }，C z The actual residual capacity of the lithium battery after the Z-th charge-discharge cycle is represented, and Z represents the total charge-discharge cycle number of the lithium battery;
step two, health factor data p of the z-th charge-discharge cycle after pretreatment z Performing pearson correlation coefficient analysis on any two health factors, and adding a pair of health factors with pearson correlation coefficient scores larger than a set threshold value into a health factor data set F with strong correlation;
step three, constructing a convolution time sequence attention network model M, which comprises the following steps: the device comprises a feature extraction module, a feature fusion module and a fusion attention prediction module; and an exponential linear function ReLU is adopted as an activation function of a convolution time sequence attention network model M;
step 3.1, the feature extraction module consists of two one-dimensional convolution layers and a maximum pooling layer, and processes the charge and discharge data sets with uniform data length under the z-th charge and discharge cycle to obtain a feature sequence F of the charge and discharge data sets under the z-th charge and discharge cycle z ＝{F Uz ，F Tz ，F Iz And get the feature map q= { F 1 ,F 2 ,…,F z …,F Z -a }; wherein F is Uz Represents the discharge voltage U at the z-th charge-discharge cycle z And F Uz ＝{f U1 ,f U2 …,f Uk …f UK }，f Uk Represents F Uz The kth feature value, F in the feature sequence Tz Indicating the battery temperature T at the z-th charge-discharge cycle z Characteristic sequence of F (F) Tz ＝{f T1 ,f T2 …,f Tk …f TK }；f Tk Represents F Tz The kth feature value, F in the feature sequence Iz Indicating the charge current at the z-th charge-discharge cycleI z Is a characteristic sequence of (2); f (F) Iz ＝{f I1 ,f I2 …,f Ik …f IK }；f Ik Representation of F Iz The K-th characteristic value in the characteristic sequence, wherein K represents the characteristic number output after passing through the characteristic extraction module;
step 3.2, the characteristic fusion module splices the characteristic map Q with voltage, current and temperature data in the health factor data set F according to characteristic classification according to the condition of each charge-discharge cycle, so as to obtain a fusion characteristic map Q';
step 3.3, the fused attention prediction module consists of two layers of PLSTM networks and a time sequence attention module, wherein each layer of PLSTM network consists of Z memory cells, and the time sequence attention module consists of a convolution layer, an attention layer and a full connection layer;
step 3.3.1, after the fusion feature map Q' is processed through the first layer pls network, obtaining a hidden layer state h= { H 1 ,h 2 …,h z …h Z And (b) wherein h z Representing the hidden layer state of the z-th memory cell output;
step 3.3.2, after the hidden layer state H is processed by a layer of convolution layer, obtaining the convolution characteristic of the hidden layer
Wherein (1)>
Representing the z-th convolution characteristic after convolution;
step 3.3.3, the attention layer convolves the hidden layer state H and the hidden layer convolution feature H with the use of (2) C Calculating to obtain an attention weight set A= { A 1 ,A 2 …,A z …A Z (wherein A) z Represents the z-th attention score;
in the formula (2), W a Representing a parameter matrix to be learned; t represents a transpose;
step 3.3.4, after the attention weight set A is spliced with the hidden layer state H output by the PLSTM network through the step 3, the attention fusion characteristic H' = { H is obtained 1 ′,h 2 ′…,h z ′…h Z ' s; wherein h is z ' represents the z-th attention fusion feature;
in the formula (3), alpha and beta are two super parameters, h r Representing the state value of the r hidden layer;
step 3.3.5, inputting H' into the second layer PLSTM network for processing, thereby obtaining a predicted sequence Y= { Y 1 ,y 2 …,y z …y Z -a }; wherein y is z Representing the predicted value output by the Z-th memory cell of the second layer PLS (PLS) network, and enabling the predicted value y output by the Z-th memory cell of the second layer PLS (TM) network Z A predicted value for the final lithium battery capacity value;
step 3.3.6, converting the predicted sequence Y into a final SOH value by using a formula (4);
in the formula (4), C 0 Representing the rated capacity of a lithium battery, SOH z Representing a battery state of health value at a z-th charge-discharge cycle;
training a convolution time sequence attention network model M;
step 4.1, constructing a mean square error Loss function Loss by using the method (5) MSE (y z ,C z )；
In the formula (5), C z A true value representing the capacity of the lithium battery at the z-th charge-discharge cycle, y z A predicted value output by the z-th memory cell;
step 4.2, optimizing all parameters in the convolution time sequence attention network model M by adopting an Adam optimizer until the total Loss is low MSE No further descent; the optimized space-time attention network model is obtained and used for predicting the health state of the lithium battery.
The invention provides an electronic device, which comprises a memory and a processor, and is characterized in that the memory is used for storing a program for supporting the processor to execute the lithium battery health state estimation method, and the processor is configured to execute the program stored in the memory.
The invention relates to a computer readable storage medium, which is stored with a computer program, characterized in that the computer program when being run by a processor executes the steps of the lithium battery health state estimation method.
Compared with the prior art, the invention has the beneficial effects that:
1. the invention optimizes the input data characteristics by using pearson correlation coefficient analysis and characteristic extraction technology, and selects the training data which is most favorable for model training. A plurality of health factors are used for training, uncertainty of training of single health factor is reduced, accuracy of estimating health state of the lithium battery by the model is improved, high-dimensional characteristics between single charge and discharge are learned by a convolution method, and the problem that data in a learning cycle cannot be accurately linked is solved.
2. According to the invention, the PLSTM network is adopted, so that the snoop hole structure is increased, each gate layer of the LSTM also receives the input of the cell state, once new information is adopted, information related to preference memory can be extracted, and unnecessary error signals are shielded, so that more beneficial information can be screened, and the extraction efficiency of the traditional LSTM network on information between charge and discharge cycles is improved.
3. According to the invention, the output of the first PLSTM network is input into the time sequence attention module so as to capture the hierarchical characteristics among several variables affecting the battery capacity degradation and the time dependence of embedding in the characteristics, so that the long-term dependence of lithium battery data is better learned, key data are found out through attention weight searching to make a mat for PLSTM network learning of the next layer, and finally, the data representation in charge-discharge cycles and between cycles can be extracted efficiently and accurately, and the accuracy of lithium battery health state prediction is remarkably improved.
Drawings
FIG. 1 is a flow chart of lithium battery state of health estimation according to the present invention;
FIG. 2 is a graph of the partial cycling discharge voltage of a lithium battery in accordance with the present invention;
FIG. 3 is a graph of the charge current of a lithium battery in accordance with the present invention for a partial cycle;
FIG. 4 is a graph of the temperature at discharge for a partial cycle of a lithium battery in accordance with the present invention;
FIG. 5 is a block diagram of a SOH estimation model of a lithium battery based on CNN-PLSTM;
FIG. 6 is a diagram showing the structure of PLSTM model memory cells according to the present invention;
fig. 7 is a timing attention pattern diagram of the present invention.
Detailed Description
In this embodiment, a method for estimating a health state of a lithium battery based on feature selection and time sequence attention is to fully utilize data feature information contained in a charge-discharge period and extract multiple health factors and features in the period into a model. Meanwhile, considering the long-term dependence of time sequence information, a time sequence attention module is designed for enhancing the characteristic expression and important data information during the charge and discharge period; and secondly, in order to enhance the study of the long-term and short-term information of the model, a PLSTM network is added into the model for enhancing the extraction of LSTM cells on long-term and short-term memory. The overall flow of the method is shown in fig. 1, and the method comprises the following steps: acquiring lithium battery data with label labels under experimental conditions, and preprocessing to realize extraction of health factors of the lithium battery data and feature extraction during charge and discharge cycles; then, the two parts of features are fused through a feature fusion module, and a global feature map with higher data representation is obtained; thirdly, enabling the model to learn the long-term dependence of data and the internal connection of each characteristic between charge and discharge periods through a PLSTM network and a time sequence attention mechanism; and finally, outputting the lithium battery capacity predicted by the model and converting the lithium battery capacity into a final predicted SOH value according to a related formula. The method can effectively improve the accuracy of lithium battery health state prediction, is simple in model realization, easy in acquisition of training data, has the characteristics of high practicality, high prediction accuracy and the like, and is suitable for online lithium battery health state prediction and lithium battery capacity estimation:
step one, acquiring data sets such as NASA, mendeliy, CALCE of battery capacity, voltage, temperature and current of different rounds in multiple charge and discharge cycles of a lithium battery, and preprocessing;
step 1.1, acquiring discharge voltage, charge current and discharge temperature data of a lithium battery under a z-th charge-discharge cycle, performing abnormal value removal treatment, and respectively performing uniform data length treatment on the discharge voltage, charge current and discharge temperature data of the lithium battery after abnormal values are removed to obtain a charge-discharge data set with uniform data length under the z-th charge-discharge cycle, wherein the recorded number length of each charge-discharge cycle in the lithium battery data set is inconsistent, so that the data of each charge-discharge cycle needs to be treated according to a time span, thereby facilitating the next operation; comprising the following steps: discharge voltage U z ＝{u z1 ,u z2 ,u z3 …,u zj …,u zJ As shown in fig. 2, as the number of charge and discharge increases, the time until the voltage reaches the maximum value becomes earlier, and the duration of discharge becomes shorter. Charging current data I z ＝{i z1 ,i z2 ,…,i zj …,i zJ As shown in fig. 3, as the number of charge and discharge increases, the charge time becomes shorter, indicating that the capacity of the battery is gradually decreasing; charge-discharge temperature data T z ＝{t z1 ,t z2 ,t z3 …,t zj …,t zJ -a }; as shown in fig. 4, the charge and discharge times are increasedThe time for the temperature to reach the peak is earlier and the temperature value is higher and higher; so the training using the related voltage, current and temperature data has the related facts basis; wherein u is zj Representing a jth discharge voltage value in a zth discharge cycle; i.e zj Representing a jth charging current value in a z-th charging cycle; t is t zj Representing a jth discharge temperature value in a zth discharge cycle; j represents the record times of one charge-discharge cycle in the charge-discharge data set; where j=200
Step 1.2, respectively extracting maximum value and minimum value data of discharge voltage, charge current and discharge temperature data of the lithium battery in a z-th charge-discharge cycle, and carrying out averaging operation to obtain nine health factor data consisting of the maximum value, the minimum value and the average value of the charge-discharge data of the lithium battery in the z-th charge-discharge cycle, wherein the method comprises the following steps: voltage health factor U z ′＝{u z,max ,u z,min ,u z,avg -a }; current health factor I z ′＝{i z,max ,i z,min ,i z,avg -a }; temperature health factor T z ′＝{t z,max ,t z,min ,t z,avg }, where u z,max Representing the maximum value of the discharge voltage in the z-th discharge cycle; u (u) z,min Representing the minimum value of the discharge voltage in the z-th discharge cycle; u (u) z,avg Representing the average value of the discharge voltage in the z-th discharge cycle; i.e z,max Representing the maximum value of the charging current in the z-th charging cycle; i.e z,min Representing a minimum value of the charging current in the z-th charging cycle; i.e z,avg Representing an average value of the charging current in the z-th charging cycle; t is t z,max Representing the maximum value of the battery temperature in the z-th discharge cycle; t is t z,min Representing a minimum value of the battery temperature in the z-th discharge cycle; t is t z,avg Mean value of battery temperature in z-th discharge cycle;
step 1.3, splicing nine health factor data with the capacity of the lithium battery, and performing normalization processing to obtain preprocessed health factor data P= { P 1 ,p 2 ,…,p z …,p Z P is }, where z Indicating the z-th charge after pretreatmentHealth factor data of discharge cycle, and p z ＝{U z ′,I z ′,T z ′,C z }，C z The actual residual capacity of the lithium battery after the Z-th charge-discharge cycle is represented, and Z represents the total charge-discharge cycle number of the lithium battery; wherein z=168;
step two, health factor data p of the z-th charge-discharge cycle after pretreatment z Performing pearson correlation coefficient analysis on any two health factors, and adding a pair of health factors with pearson correlation coefficient scores larger than a set threshold value into a health factor data set F with strong correlation; the features of the input data can be optimized through pearson correlation coefficient analysis, training data which is most favorable for model training is selected, useless data input is reduced, and the model training speed is accelerated;
step three, constructing a convolution time sequence attention network model M as shown in fig. 5, which comprises the following steps: the device comprises a feature extraction module, a feature fusion module and a fusion attention prediction module; and an exponential linear function ReLU is adopted as an activation function of a convolution time sequence attention network model M; the fusion attention prediction module captures hierarchical features among several variables affecting battery capacity degradation and time dependence embedded in the features through a PLSTM network, and gives attention weight obtained by using time sequence attention to weight information among each charge and discharge cycle, so that the network pays attention to learning key data, and interference of invalid information is reduced;
step 3.1, the feature extraction module consists of two one-dimensional convolution layers and a maximum pooling layer, and processes the charge and discharge data sets with uniform data length under the z-th charge and discharge cycle to obtain a feature sequence F of the charge and discharge data set under the z-th charge and discharge cycle z ＝{F Uz ，F Tz ，F Iz And get the feature map q= { F 1 ,F 2 ,…,F z …,F z -a }; wherein F is Uz Represents the discharge voltage U at the z-th charge-discharge cycle z And F Uz ＝{f U1 ,f U2 …,f Uk …f UK }，f Uk Represents F Uz The k-th feature value in the feature sequence,F Tz indicating the battery temperature T at the z-th charge-discharge cycle z Characteristic sequence of F (F) Tz ＝{f T1 ,f T2 ...,f Tk ...f Tk }；f Tk Represents F Tz The kth feature value, F in the feature sequence Iz Indicating the charge current I under the z-th charge-discharge cycle z Is a characteristic sequence of (2); f (F) Iz ＝{f I1 ,f I2 ...,f Ik …f IK }；F Ik Representation of F Iz The K-th characteristic value in the characteristic sequence, wherein K represents the characteristic number output after passing through the characteristic extraction module; wherein k=20
Step 3.2, the characteristic fusion module splices the characteristic diagram Q with voltage, current and temperature data in the health factor data set F according to characteristic classification according to the condition of each charge-discharge cycle, so as to obtain a fusion characteristic diagram Q';
step 3.3, a fused attention prediction module consists of two layers of PLSTM networks and a time sequence attention module, wherein each layer of PLSTM network consists of Z memory cells, and the time sequence attention module consists of a convolution layer, an attention layer and a full connection layer;
the PLSTM network memory cell is shown in figure 6, and comprises three gates, namely a forgetting gate, an input gate and an output gate, wherein the forgetting gate can be used for selectively forgetting information in the cell state of the last step, the input gate can be used for selectively recording new information into the cell state, and the output gate can be used for determining output prediction information; meanwhile, a snoop hole structure is added, so that each door of PLSTM also receives the input of the cell state, once new information is passed, information related to preference memory can be extracted, and unnecessary error signals are shielded, so that more beneficial information can be screened out;
the time sequence attention module is shown in fig. 7, firstly, a fixed-length time sequence mode in input information is extracted by convolution, then a scoring function is set to be that the convolved information and the original hidden layer information are subjected to matrix multiplication to carry out attention weighting, then the output of the attention score is normalized to be between 0 and 1 through a Sigmoid function, and finally, the final attention fusion characteristic is obtained by weighting with the hidden layer state of each memory cell; thereby creating an attention mechanism over the timing of the features,
step 3.3.1, after the fusion feature map Q' is processed through the first layer pls network, obtaining a hidden layer state h= { H 1 ,h 2 …,h z …h Z And (b) wherein h z Representing the hidden layer state of the z-th memory cell output;
step 3.3.2, after the hidden layer state H is processed by a layer of convolution layer, obtaining the convolution characteristic of the hidden layer
Wherein (1)>
Representing the z-th convolution characteristic after convolution;
step 3.3.3, attention layer utilizing (2) Condition H of hidden layer and hidden layer convolution feature H C Calculating to obtain an attention weight set A= { A 1 ,A 2 …,A z …A Z (wherein A) z Represents the z-th attention score;
in the formula (2), W a Representing a parameter matrix to be learned; t represents a transpose;
step 3.3.4, after the attention weight set A is spliced with the hidden layer state H output by the PLSTM network through the step 3, the attention fusion characteristic H' = { H is obtained 1 ′,h 2 ′…,h z ′…h Z ' s; wherein h is z ' represents the z-th attention fusion feature;
in the formula (3), alpha and beta are two super parameters, h r Representing the state value of the r hidden layer; due to the differences in the contribution of the attention weight and the hidden layer weight to the model and their numerical ranges, a simple addition may result in numerical distortion that results in a model that cannot be fitted. Therefore, the weights of the two loss functions are adjusted by adding alpha and beta to the formula, and the values are respectively set to be 0.3 and 0.7 in the experiment
Step 3.3.5, inputting H' into the second layer PLSTM network for processing, thereby obtaining a predicted sequence Y= { Y 1 ,y 2 …,y z …y Z -a }; wherein y is z Representing the predicted value output by the Z-th memory cell of the second layer PLS (PLS) network, and enabling the predicted value y output by the Z-th memory cell of the second layer PLS (TM) network Z A predicted value for the final lithium battery capacity value; because the PLSTM network model is a single-step prediction model, that is, the next charge-discharge cycle data is predicted through the input sequence, the last bit of the prediction sequence is output as the predicted lithium battery capacity value;
step 3.3.6, converting the predicted sequence Y into a final SOH value by using a formula (4);
in the formula (4), C 0 Representing the rated capacity of a lithium battery, SOH z Representing a battery state of health value at a z-th charge-discharge cycle;
training a convolution time sequence attention network model M;
step 4.1, constructing a mean square error Loss function Loss by using the method (5) MSE (y z ,C z )；
In the formula (5), C z A true value representing the capacity of the lithium battery at the z-th charge-discharge cycle, y z Predicted value of z-th memory cell output
Step 4.2, optimizing all parameters in the convolution time sequence attention network model M by adopting an Adam optimizer, wherein the learning rate is set to be 0.001, the batch size is set to be 32 by a sliding window method, and the total iteration times are 1500 times until the total Loss is achieved MSE No further descent; the optimized space-time attention network model is obtained and used for predicting the health state of the lithium battery.
In this embodiment, an electronic device includes a memory for storing a program for supporting the processor to execute the above lithium battery health state estimation method, and a processor configured to execute the program stored in the memory.
In this embodiment, a computer readable storage medium stores a computer program, which when executed by a processor, performs the steps of the lithium battery health state estimation method described above.
Claims (3)
1. A lithium battery health state estimation method based on feature selection and time sequence attention is characterized by comprising the following steps:
step one, acquiring data sets of battery capacity, voltage, temperature and current of different rounds in multiple charge and discharge cycles of a lithium battery, and preprocessing the data sets;
step 1.1, obtain discharge voltage, charge current and discharge temperature data of lithium battery under the z-th charge-discharge cycle and remove abnormal value processing, then carry on the processing of unifying the data length respectively with lithium battery discharge voltage, charge current and discharge temperature data after removing the abnormal value, get the unified charge-discharge dataset of data length under the z-th charge-discharge cycle, including: discharge voltage U z ＝{u z1 ，u z2 ，u z3 …，u zj …，u zJ Charge current data I z ＝{i z1 ，i z2 ，…，i zj …，i zJ Charge-discharge temperature data T z ＝{t z1 ，t z2 ，t z3 …，t zj …，t zJ -a }; wherein u is zj Representing a jth discharge voltage value in a zth discharge cycle; i.e zj Representing a jth charging current value in a z-th charging cycle; t is t zj Representing a jth discharge temperature value in a zth discharge cycle; j represents the record times of one charge-discharge cycle in the charge-discharge data set;
step 1.2, respectively extracting maximum value and minimum value data of discharge voltage, charge current and discharge temperature data of the lithium battery in a z-th charge-discharge cycle, and carrying out averaging operation to obtain nine health factor data consisting of the maximum value, the minimum value and the average value of the charge-discharge data of the lithium battery in the z-th charge-discharge cycle, wherein the method comprises the following steps: voltage health factor U z ′＝{u z，max ，u z，min ，u z，avg -a }; current health factor I z ′＝{i z，max ，i z，min ，i z，avg -a }; temperature health factor T z ′＝{t z，max ，t z，min ，t z，avg }, where u z，max Representing the maximum value of the discharge voltage in the z-th discharge cycle; u (u) z，min Representing the minimum value of the discharge voltage in the z-th discharge cycle; u (u) z，avg Representing the average value of the discharge voltage in the z-th discharge cycle; i.e z，max Representing the maximum value of the charging current in the z-th charging cycle; i.e z，min Representing a minimum value of the charging current in the z-th charging cycle; i.e z，avg Representing an average value of the charging current in the z-th charging cycle; t is t z，max Representing the maximum value of the battery temperature in the z-th discharge cycle; t is t z，min Representing a minimum value of the battery temperature in the z-th discharge cycle; t is t z，avg Mean value of battery temperature in z-th discharge cycle;
step 1.3, splicing nine health factor data with the capacity of the lithium battery, and performing normalization processing to obtain preprocessed health factor data P= { P 1 ，p 2 ，…，p z …，p Z P is }, where z Health factor data representing the z-th charge-discharge cycle after pretreatment, and p z ＝{U z ′，I z ′，T z ′，C z }，C z Indicating z < thThe actual residual capacity of the lithium battery after the secondary charge and discharge cycles, Z represents the total charge and discharge cycle number of the lithium battery;
step two, health factor data p of the z-th charge-discharge cycle after pretreatment z Performing pearson correlation coefficient analysis on any two health factors, and adding a pair of health factors with pearson correlation coefficient scores larger than a set threshold value into a health factor data set F with strong correlation;
step three, constructing a convolution time sequence attention network model M, which comprises the following steps: the device comprises a feature extraction module, a feature fusion module and a fusion attention prediction module; and an exponential linear function ReLU is adopted as an activation function of a convolution time sequence attention network model M;
step 3.1, the feature extraction module consists of two one-dimensional convolution layers and a maximum pooling layer, and processes the charge and discharge data sets with uniform data length under the z-th charge and discharge cycle to obtain a feature sequence F of the charge and discharge data sets under the z-th charge and discharge cycle z ＝{F Uz ，F Tz ，F Iz And get the feature map q= { F 1 ，F 2 ，…，F z …，F Z -a }; wherein F is Uz Represents the discharge voltage U at the z-th charge-discharge cycle z And F Uz ＝{f U1 ，f U2 …，f Uk …f UK }，f Uk Represents F Uz The kth feature value, F in the feature sequence Tz Indicating the battery temperature T at the z-th charge-discharge cycle z Characteristic sequence of F (F) Tz ＝{f T1 ，f T2 …，f Tk …f TK }；f Tk Represents F Tz The kth feature value, F in the feature sequence Iz Indicating the charge current I under the z-th charge-discharge cycle z Is a characteristic sequence of (2); f (F) Iz ＝{f I1 ，f I2 …，f Ik …f IK }；f Ik Representation of F Iz The K-th characteristic value in the characteristic sequence, wherein K represents the characteristic number output after passing through the characteristic extraction module;
step 3.2, the characteristic fusion module splices the characteristic map Q with voltage, current and temperature data in the health factor data set F according to characteristic classification according to the condition of each charge-discharge cycle, so as to obtain a fusion characteristic map Q';
step 3.3, the fused attention prediction module consists of two layers of PLSTM networks and a time sequence attention module, wherein each layer of PLSTM network consists of Z memory cells, and the time sequence attention module consists of a convolution layer, an attention layer and a full connection layer;
step 3.3.1, after the fusion feature map Q' is processed through the first layer pls network, obtaining a hidden layer state h= { H 1 ，h 2 …，h z …h Z And (b) wherein h z Representing the hidden layer state of the z-th memory cell output;
step 3.3.2, after the hidden layer state H is processed by a layer of convolution layer, obtaining the convolution characteristic of the hidden layer
Wherein (1)>
Representing the z-th convolution characteristic after convolution;
step 3.3.3, the attention layer convolves the hidden layer state H and the hidden layer convolution feature H with the use of (2) C Calculating to obtain an attention weight set A= { A 1 ，A 2 …，A z …A Z (wherein A) z Represents the z-th attention score;
in the formula (2), W a Representing a parameter matrix to be learned; t represents a transpose;
step 3.3.4, the attention weight set A and the hidden output by a layer of PLSTM network are outputted through the step (3)After the hidden layer state H is spliced, attention fusion characteristics H' = { H are obtained 1 ′，h 2 ′…，h z ′…h Z ' s; wherein h is z ' represents the z-th attention fusion feature;
in the formula (3), alpha and beta are two super parameters, h r Representing the state value of the r hidden layer;
step 3.3.5, inputting H' into the second layer PLSTM network for processing, thereby obtaining a predicted sequence Y= { Y 1 ，y 2 …，y z …y Z -a }; wherein y is z Representing the predicted value output by the Z-th memory cell of the second layer PLS (PLS) network, and enabling the predicted value y output by the Z-th memory cell of the second layer PLS (TM) network Z A predicted value for the final lithium battery capacity value;
step 3.3.6, converting the predicted sequence Y into a final SOH value by using a formula (4);
in the formula (4), C 0 Representing the rated capacity of a lithium battery, SOH z Representing a battery state of health value at a z-th charge-discharge cycle;
training a convolution time sequence attention network model M;
step 4.1, constructing a mean square error Loss function Loss by using the method (5) MSE (y z ，C z )；
In the formula (5), C z A true value representing the capacity of the lithium battery at the z-th charge-discharge cycle, y z A predicted value output by the z-th memory cell;
step 4.2, optimizing all parameters in the convolution time sequence attention network model M by adopting an Adam optimizer until the total Loss is low MSE No further descent; the optimized space-time attention network model is obtained and used for predicting the health state of the lithium battery.
2. An electronic device comprising a memory and a processor, wherein the memory is configured to store a program that supports the processor to perform the lithium battery state of health estimation method of claim 1, the processor being configured to execute the program stored in the memory.
3. A computer readable storage medium having a computer program stored thereon, characterized in that the computer program when run by a processor performs the steps of the lithium battery state of health estimation method of claim 1.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310082075.8A CN116068399A (en) | 2023-02-08 | 2023-02-08 | Lithium battery health state estimation method based on feature selection and time sequence attention |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310082075.8A CN116068399A (en) | 2023-02-08 | 2023-02-08 | Lithium battery health state estimation method based on feature selection and time sequence attention |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN116068399A true CN116068399A (en) | 2023-05-05 |
Family
ID=86169558
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202310082075.8A Pending CN116068399A (en) | 2023-02-08 | 2023-02-08 | Lithium battery health state estimation method based on feature selection and time sequence attention |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN116068399A (en) |
Cited By (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116298934A (en) * | 2023-05-19 | 2023-06-23 | 河南科技学院 | Modeling method of prediction network for lithium battery health state estimation |
| CN116609676A (en) * | 2023-07-14 | 2023-08-18 | 深圳先进储能材料国家工程研究中心有限公司 | Method and system for monitoring state of hybrid energy storage battery based on big data processing |
| CN116840700A (en) * | 2023-08-31 | 2023-10-03 | 深圳市安德普电源科技有限公司 | Method, device, equipment and storage medium for monitoring battery state in real time |
| CN117233615A (en) * | 2023-11-10 | 2023-12-15 | 中油绿电新能源有限公司 | Battery charging process abnormality detection method and device based on comparison learning network |
| CN117269766A (en) * | 2023-08-28 | 2023-12-22 | 广东工业大学 | Battery SOH prediction method for unbalanced use scene |
| CN117590260A (en) * | 2024-01-18 | 2024-02-23 | 武汉船用电力推进装置研究所(中国船舶集团有限公司第七一二研究所) | Method and device for estimating state of charge of marine lithium ion power battery and electronic equipment |
| CN117665630A (en) * | 2024-02-01 | 2024-03-08 | 云储新能源科技有限公司 | Battery life prediction method and system based on charge-discharge cycle data |
-
2023
- 2023-02-08 CN CN202310082075.8A patent/CN116068399A/en active Pending
Cited By (13)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116298934B (en) * | 2023-05-19 | 2023-08-04 | 河南科技学院 | Modeling method of prediction network for lithium battery health state estimation |
| CN116298934A (en) * | 2023-05-19 | 2023-06-23 | 河南科技学院 | Modeling method of prediction network for lithium battery health state estimation |
| CN116609676A (en) * | 2023-07-14 | 2023-08-18 | 深圳先进储能材料国家工程研究中心有限公司 | Method and system for monitoring state of hybrid energy storage battery based on big data processing |
| CN116609676B (en) * | 2023-07-14 | 2023-09-15 | 深圳先进储能材料国家工程研究中心有限公司 | Method and system for monitoring state of hybrid energy storage battery based on big data processing |
| CN117269766A (en) * | 2023-08-28 | 2023-12-22 | 广东工业大学 | Battery SOH prediction method for unbalanced use scene |
| CN116840700A (en) * | 2023-08-31 | 2023-10-03 | 深圳市安德普电源科技有限公司 | Method, device, equipment and storage medium for monitoring battery state in real time |
| CN116840700B (en) * | 2023-08-31 | 2023-10-31 | 深圳市安德普电源科技有限公司 | Method, device, equipment and storage medium for monitoring battery state in real time |
| CN117233615A (en) * | 2023-11-10 | 2023-12-15 | 中油绿电新能源有限公司 | Battery charging process abnormality detection method and device based on comparison learning network |
| CN117233615B (en) * | 2023-11-10 | 2024-02-06 | 中油绿电新能源有限公司 | Battery charging process abnormality detection method and device based on comparison learning network |
| CN117590260A (en) * | 2024-01-18 | 2024-02-23 | 武汉船用电力推进装置研究所(中国船舶集团有限公司第七一二研究所) | Method and device for estimating state of charge of marine lithium ion power battery and electronic equipment |
| CN117590260B (en) * | 2024-01-18 | 2024-04-16 | 武汉船用电力推进装置研究所(中国船舶集团有限公司第七一二研究所) | Method and device for estimating state of charge of marine lithium ion power battery and electronic equipment |
| CN117665630A (en) * | 2024-02-01 | 2024-03-08 | 云储新能源科技有限公司 | Battery life prediction method and system based on charge-discharge cycle data |
| CN117665630B (en) * | 2024-02-01 | 2024-04-12 | 云储新能源科技有限公司 | Battery life prediction method and system based on charge-discharge cycle data |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN116068399A (en) | Lithium battery health state estimation method based on feature selection and time sequence attention | |
| CN112782591B (en) | Lithium battery SOH long-term prediction method based on multi-battery data fusion | |
| CN112798960B (en) | Battery pack residual life prediction method based on migration deep learning | |
| CN113009349A (en) | Lithium ion battery health state diagnosis method based on deep learning model | |
| CN111832221B (en) | Lithium battery life prediction method based on feature screening | |
| CN113917334B (en) | Battery health state estimation method based on evolution LSTM self-encoder | |
| Yayan et al. | A novel method for SoH prediction of batteries based on stacked LSTM with quick charge data | |
| CN111983474A (en) | Lithium ion battery life prediction method and system based on capacity decline model | |
| CN112098873B (en) | Lithium battery health state estimation method based on charging voltage curve geometric characteristics | |
| CN112686380A (en) | Neural network-based echelon power cell consistency evaluation method and system | |
| CN115656857A (en) | Multi-scale fusion prediction method for remaining service life of lithium ion battery | |
| CN115980584A (en) | Lithium battery RUL estimation method based on multi-feature fusion LSTM network | |
| CN115308608A (en) | All-vanadium redox flow battery voltage prediction method, device and medium | |
| CN116699410A (en) | Lithium battery health state estimation method integrating unsupervised and supervised learning | |
| CN114545275A (en) | Indirect prediction method for remaining service life of lithium ion battery | |
| CN113657030A (en) | Method for predicting remaining service life of lithium battery based on Gaussian process regression | |
| CN112881914B (en) | Lithium battery health state prediction method | |
| CN116400246A (en) | ACO-GMR-based energy storage power station battery health state estimation method | |
| CN114280491B (en) | Retired battery residual capacity estimation method based on active learning | |
| CN115481796A (en) | Method for predicting remaining service life of battery based on Bayesian hybrid neural network | |
| Lin et al. | Li-ion battery state of health prediction based on long short-term memory recurrent neural network | |
| CN113255215B (en) | Lithium battery health state estimation method based on voltage segments | |
| CN113093014B (en) | Online collaborative estimation method and system for SOH and SOC based on impedance parameters | |
| CN115236521A (en) | Lithium ion battery health state estimation method based on semi-supervised cooperative training framework | |
| CN115372829A (en) | BMS data-based method for rapidly evaluating health degree of vehicle battery |
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 |