Lithium battery capacity prediction method based on empirical mode decomposition and data fusion

Abstract

The lithium battery capacity prediction method based on empirical mode decomposition and data fusion drive provided by the present invention includes step S1, obtaining the historical capacity degradation data of a lithium battery to form a source data set; step S2, iteratively screening the source data set using empirical mode decomposition, and decomposing its historical capacity degradation data into a residual sequence and multiple intrinsic mode functions; step S3, constructing a long short-term memory neural network source prediction model, and pre-training the constructed source prediction model using the residual sequence and intrinsic mode function data obtained in step S2; step S4, fine-tuning the parameters of the pre-trained source prediction model to achieve transfer learning and obtain the optimal prediction model; step S5, predicting the future capacity of the target lithium battery using the optimal prediction model obtained in step S4. The present invention can effectively predict the future capacity and remaining life of lithium batteries.

Description

Technical Field

[0001] The present invention relates to the field of energy control technology, and in particular to a lithium battery capacity prediction method based on empirical mode decomposition and data fusion drive. Background Art

[0002] As the primary or secondary energy source of many power systems, batteries undergo continuous charge and discharge cycles, and the electrochemical reaction inside the battery slows down, causing degradation of the electrode material, resulting in reduced battery capacity and increased internal resistance. In severe cases, serious faults such as internal short circuits may occur, causing serious consequences. Therefore, advance knowledge of the relevant battery status information can effectively avoid safety hazards caused by lithium-ion battery aging and performance degradation.

[0003] Currently, most methods for estimating the future capacity and remaining life of batteries use model or data-driven methods.

[0004] The model-driven approach is a method that analyzes the internal parameters of the battery by establishing a mathematical or physical model to estimate how much time or number of cycles the battery will have remaining in future use. Common methods include those based on equivalent circuit models, electrochemical models, and empirical degradation models. The equivalent circuit model is an empirical model that does not describe the complex structure inside the battery and the uncertainty during operation, so its prediction accuracy needs to be considered. The electrochemical model strictly describes the electrochemical reactions inside the lithium battery, but its model complexity is high, the number of model parameters is huge, and the identification is difficult, which limits its scope of application.

[0005] Based on the data-driven method, by extracting characteristic parameters related to battery life from a large amount of raw data, an appropriate machine learning model is selected to establish the corresponding relationship between input and remaining capacity, so as to predict the future capacity of the battery. Among them, the more widely used machine learning models include artificial neural network (ANN), support vector machine (SVM) and relevance vector machine (RVM). In the data-driven method, the uncertainty of a large amount of data and the uncertainty of the parameters of the single data-driven model often affect the final prediction performance. Secondly, the degradation of lithium batteries is not completely consistent when facing different stresses, and the degradation curve is easily affected by other factors, which makes the established model less generalizable, thus limiting the prediction scenario of the model. Summary of the invention

[0006] In order to improve the accuracy of the prediction of the future capacity of a lithium battery, the present invention provides a lithium battery capacity prediction method based on empirical mode decomposition and data fusion drive.

[0007] In order to solve the above technical problems, the present invention adopts the following technical method: a lithium battery capacity prediction method based on empirical mode decomposition and data fusion drive, comprising:

[0008] Step S1, obtaining historical capacity degradation data of a lithium battery to form a source data set;

[0009] Step S2, using empirical mode decomposition to iteratively screen the source data set, decomposing its historical capacity degradation data into a residual sequence and multiple intrinsic mode functions, wherein the high-frequency parts of capacity regeneration and capacity fluctuation are represented by the intrinsic mode function part, and the overall trend of capacity degradation is represented by the residual sequence;

[0010] Step S3, constructing a long short-term memory neural network source prediction model, and pre-training the constructed source prediction model using the residual sequence and intrinsic mode function data obtained in step S2;

[0011] The source prediction model includes an input layer, a hidden layer and an output layer connected in sequence; wherein the number of neurons in the input layer is the sum of the number of intrinsic mode functions and residual sequences obtained after the original capacity degradation data is subjected to empirical mode decomposition; the hidden layer includes two LSTM layers connected in sequence and a dense connection layer, the two LSTM layers learn the common characteristics of lithium battery capacity degradation data, consider the correlation between capacity degradation time series data through effective long-term dependency learning, and the dense connection layer performs transfer learning on the source prediction model to fine-tune the model; the number of outputs in the output layer is the number of output categories of the target task;

[0012] Step S4, fine-tuning the parameters of the pre-trained source prediction model to implement transfer learning and obtain the optimal prediction model;

[0013] First, the historical capacity degradation data of the target lithium battery is obtained to form a target domain data set. Then, the target domain data set is processed according to step S2 to obtain input data. Then, two LSTM layers of the source prediction model are directly copied. Then, the dense connection layer and the output layer of the source prediction model are added to the copied LSTM layer through random parameter initialization. Finally, the parameters of the copied LSTM layer are fine-tuned based on the source prediction model using the input data. The dense connection layer and the output layer connected to the copied LSTM layer are trained from scratch to obtain the optimal prediction model.

[0014] Step S5, using the optimal prediction model obtained in step S4 to predict the future capacity of the target lithium battery.

[0015] Furthermore, in step S1, historical capacity degradation data of the lithium battery is obtained by monitoring the capacity of the lithium battery at different charge and discharge cycle stages.

[0016] Furthermore, the step S2 includes:

[0017] S201, the historical capacity degradation data of lithium batteries After normalization, we get :

[0018] (1)

[0019] in, and They are the maximum and minimum values ​​in the historical capacity degradation data of lithium batteries;

[0020] S202, using empirical mode decomposition to analyze the data Perform iterative screening, from the number of iterations p =1 to the maximum number of iterations , do the following:

[0021] 1) Select Settings Data , by comparing adjacent values ​​within all fluctuations, searching All local maxima and minima in the , and then use splines to connect these local extreme values ​​to construct the upper envelope and lower envelope ;

[0022] 2) Calculate the local average :

[0023] (2)

[0024] 3) Calculate data With the local mean difference :

[0025] (3)

[0026] 4) Pass inspection Whether the standard of the eigenmode function is met is determined contains the number of eigenmode functions, if If the signal meets the standard of intrinsic mode function, it is set as IMF signal. Remove this , to obtain the corresponding residual :

[0027] (4)

[0028] 5) Repeat steps 2) to 4) until the residual is becomes a monotonic function; if the predefined maximum number of iterations is reached , the screening cycle is terminated; after this transformation process, the intrinsic mode function contains the information of capacity regeneration and capacity fluctuation, and the data It is expressed as:

[0029] (5)

[0030] in, Indicates from The number of eigenmode functions decomposed from After decomposition, The intrinsic mode functions, Represents the residuals that form a monotonic function.

[0031] Furthermore, in step S3, the future capacity predicted by the pre-trained source prediction model is expressed as follows:

[0032] (6)

[0033] in, For the Future capacity predicted by cycles; is the weight matrix of the densely connected layer, and are the weight matrices corresponding to the first LSTM layer and the second LSTM layer respectively, and are the bias vectors of the densely connected layer, the first LSTM layer, and the second LSTM layer, respectively. For the current The hidden state of the LSTM layer at this moment.

[0034] Preferably, in step S5, after the historical capacity degradation data of the target lithium battery is normalized and subjected to empirical mode decomposition, sliding time window data is constructed, the time window data size is set to 10, the current time window data is input into the optimal prediction model to predict the future capacity of the target lithium battery in the next cycle, and the predicted future capacity is updated to the window data, and a new round of sliding window data is input into the optimal prediction model to predict the future capacity of the next cycle, and this cycle is iterated. When the future capacity of the target lithium battery drops to 80% of the initial rated capacity, it is considered that the target lithium battery has reached the end of its life. At this time, the future capacity prediction curve of the target lithium battery is drawn and the remaining life of the target lithium battery is calculated, and the entire prediction process ends.

[0035] Preferably, in step S1, a B0005 18650 lithium battery with a rated capacity of 2Ah is selected, and a full charge and discharge cycle test is performed at room temperature and under constant current conditions. The test process is as follows:

[0036] 1) Discharge the lithium battery at a constant load current of 2A. When the discharge reaches the cut-off voltage of 2.6V, stop discharging and let it stand for 2h.

[0037] 2) First, charge the battery to 4.2 V with a constant current of 1.5 A, then charge it at a constant voltage of 4.2 V until the charging current is less than 0.02 A, and leave it alone for 30 min;

[0038] 3) Discharge the lithium battery at a fixed load current level of 2A until the voltage reaches 2.6 V, then stop discharging and let it stand for 30 minutes;

[0039] 4) Repeat steps 2) and 3) until the capacity of the lithium battery reaches the end of its life. After each discharge, record the discharge capacity of the battery. The discharge capacity is the actual rated capacity of the lithium battery.

[0040] In order to achieve accurate prediction of the degradation trend and uncertainty representation of lithium batteries, that is, to improve the accuracy of predicting the future capacity of lithium batteries, the present invention provides a lithium battery capacity prediction method based on empirical mode decomposition and data fusion driven, the method utilizes empirical mode decomposition EMD, long short-term memory recurrent neural network LSTM and transfer learning to learn and capture the lithium battery capacity degradation characteristics and trends from battery capacity degradation data. Among them, empirical mode decomposition is an effective signal processing technology. It has the powerful ability to extract low-frequency and high-frequency components from high-frequency dynamic signals, and can decompose the capacity degradation data of lithium-ion batteries into a series of intrinsic mode functions and residual sequences. LSTM, as a variant of recurrent neural network RNN, solves the problems of RNN gradient explosion and gradient disappearance. It is often used to deal with time series problems, and the capacity degradation data of lithium batteries belongs to time series. First, LSTM is used to learn the degradation characteristics and trends contained in the capacity degradation data, and the model is pre-trained through the source data set. The future capacity is set as the LSTM label, and a prediction of multiple groups of sequences for a single variable is realized through LSTM, so as to achieve uncertainty management of lithium-ion battery capacity degradation. Then, the model parameters are fine-tuned through a small amount of target domain data, and the trained LSTM model is transferred for learning to effectively enhance the generalization ability of the model, so that the final model can achieve effective prediction of the future capacity and remaining life of lithium batteries in the corresponding target domain. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 This is a flow chart of a lithium battery capacity prediction method based on empirical mode decomposition and data fusion drive according to the present invention;

[0042] Figure 2This is a schematic diagram comparing the decline trend of the actual rated capacity of the B0005, B0006, and B0007 lithium batteries disclosed by the NASA laboratory as the cycle period increases in this embodiment;

[0043] Figure 3 This is a schematic diagram of data obtained after normalizing the historical capacity degradation data of the B0005 18650 lithium battery in this embodiment;

[0044] Figure 4 Schematic diagram of data obtained after normalized historical capacity degradation data is subjected to empirical mode decomposition in this implementation manner;

[0045] Figure 5 Schematic diagram of the LSTM neural network unit structure of the source prediction model in the present invention;

[0046] Figure 6 A schematic diagram of obtaining an optimal prediction model by performing transfer learning on the source prediction model in the present invention;

[0047] Figure 7 This is a structural block diagram of the present invention using the optimal prediction model to predict the future capacity of a lithium battery;

[0048] Figure 8 This is a comparison chart of the predicted value and the actual value of the future capacity of target lithium battery No. B0005 in this implementation manner. DETAILED DESCRIPTION

[0049] In order to facilitate the understanding of those skilled in the art, the present invention is further described below in conjunction with embodiments and drawings. The contents mentioned in the implementation modes are not intended to limit the present invention.

[0050] The present invention predicts the future capacity and remaining life of lithium batteries based on signal processing methods and data-driven fusion methods. Since the historical capacity degradation data of lithium batteries is full of uncertainty, there will be capacity regeneration and local fluctuations, which are all high-frequency parts in the original capacity degradation data; and the lithium battery capacity degradation process shows a downward trend as a whole, and this global degradation trend is a low-frequency signal. Considering the uncertainty caused by local capacity regeneration and fluctuations during the lithium battery cycle, the historical capacity degradation data can be decomposed into low-frequency and high-frequency signals using empirical mode decomposition, and the uncertainty and global trend in the original data are considered separately to effectively improve its prediction accuracy. The LSTM model is often used to solve time series problems, and then the LSTM model is combined with transfer learning to form an LSTM+transfer learning fusion model to effectively improve the generalization ability of the model. Finally, the trained fusion model is used for iterative prediction to predict the future capacity and remaining life of the lithium battery.

[0051] like Figure 1As shown, a lithium battery capacity prediction method based on empirical mode decomposition and data fusion driving includes:

[0052] Step S1, obtaining historical capacity degradation data of a lithium battery to form a source data set.

[0053] In this embodiment, a B0005 18650 lithium battery with a rated capacity of 2Ah is selected to perform a full charge and discharge cycle test at room temperature and constant current conditions. The test process is as follows:

[0054] 1) Discharge the lithium battery at a constant load current of 2A. When the discharge reaches the cut-off voltage of 2.6V, stop discharging and let it stand for 2h.

[0055] 2) First, charge the battery to 4.2 V with a constant current of 1.5 A, then charge it at a constant voltage of 4.2 V until the charging current is less than 0.02 A, and leave it alone for 30 min;

[0056] 3) Discharge the lithium battery at a fixed load current level of 2A until the voltage reaches 2.6 V, then stop discharging and let it stand for 30 minutes;

[0057] 4) Repeat steps 2) and 3) until the capacity of the lithium battery reaches the end of its life. After each discharge, record the discharge capacity of the battery. The discharge capacity is the actual rated capacity of the lithium battery.

[0058] Since obtaining the historical capacity degradation data of the battery needs to be obtained on the basis of battery aging cycle experiments, which requires a lot of manpower and time, and the cycle test data set disclosed by NASA has been proven in many research works, the 18650 battery data of B0005 disclosed by NASA can also be selected to form the source data set required in the present invention. Taking the public source data set of NASA laboratory as an example, the actual rated capacity of lithium batteries B0005, B0006, and B0007 is compared with the decline trend of the cycle period. Figure 2 shown.

[0059] Step S2, decomposing the historical capacity degradation data of the lithium battery.

[0060] By iteratively screening the historical capacity degradation data of the source data set using empirical mode decomposition, the non-stationary source data set can be decomposed into a residual sequence and a series of modal functions (IMFs) representing orthogonal basis components. Figure 2As shown in the figure, the historical capacity decay data of lithium-ion batteries will have some fluctuations in the entire degradation process, which is expressed as the capacity regeneration and fluctuation phenomenon of the battery during the cycle. The empirical mode decomposition is used to decompose it into a residual sequence and multiple intrinsic mode functions, among which the high-frequency part of the capacity regeneration and fluctuation phenomenon can be represented by the intrinsic mode function. The intrinsic mode function quantifies the uncertainty of the lithium battery capacity degradation process, and the overall trend of capacity degradation can be represented by the residual sequence. In this way, the input of the LSTM model constructed later is no longer a single capacity data when predicting future capacity, but multiple sets of data containing uncertainty fluctuations and general trends of capacity degradation.

[0061] The screening process of decomposing historical capacity degradation data in this step is as follows:

[0062] S201, according to the following formula (1), the historical capacity degradation data of B0005 18650 lithium battery After normalization, we get Figure 3 shown Data diagram:

[0063] (1)

[0064] in, and They are respectively the maximum and minimum values ​​in the historical capacity degradation data of lithium batteries.

[0065] S202, using empirical mode decomposition to normalize the data Perform iterative screening, from the number of iterations p =1 to the maximum number of iterations , do the following:

[0066] 1) Select Settings Data , by comparing adjacent values ​​within all fluctuations, searching All local maxima and minima in the , and then use splines to connect these local extreme values ​​to construct the upper envelope and lower envelope ;

[0067] 2) Calculate the local average :

[0068] (2)

[0069] 3) Calculate data With the local mean difference :

[0070] (3)

[0071] 4) Pass inspection Whether the standard of the eigenmode function is met is determined contains the number of eigenmode functions, if If the signal meets the standard of intrinsic mode function, it is set as IMF signal. Remove this , to obtain the corresponding residual :

[0072] (4)

[0073] 5) Repeat steps 2) to 4) until the residual is becomes a monotonic function; if the predefined maximum number of iterations is reached , the screening cycle is terminated; after this transformation process, the intrinsic mode function contains the information of capacity regeneration and capacity fluctuation, and the data It is expressed as:

[0074] (5)

[0075] in, Indicates from The number of eigenmode functions decomposed from After decomposition, The intrinsic mode functions, Represents the residual that forms a monotonic function. Specifically, in this embodiment, Figure 4 As shown, the data After empirical mode decomposition, three intrinsic mode functions IMF1, IMF2, IMF3 and a residual signal are obtained.

[0076] Step S3, constructing a long short-term memory neural network source prediction model.

[0077] LSTM is an improved recurrent neural network. Compared with RNN, it optimizes the network structure and controls the neural network unit through the gate structure. Figure 5 As shown in the figure, a typical LSTM-based RNN model is divided into three gate parts: input gate, forget gate, and output gate, where:

[0078] The input gate uses sigmoid and tanh as activation functions to process the current input And the hidden state output of the previous moment , as follows:

[0079] (7)

[0080] in, is the output of the new memory unit; for Relative to The weight matrix of for Relative to The weight matrix of for The bias vector of is the output of the input gate; for for The weight matrix of for Relative to The weight matrix of for The bias vector of Represents the sigmoid activation function;

[0081] The forget gate uses sigmoid as the activation function to process the current input And the hidden state output of the previous moment , as follows:

[0082] (8)

[0083] in, is the output of the forget gate; for Relative to The weight matrix of for Relative to The weight matrix of for The bias vector of is the output of the current memory unit; is the output of the memory unit at the previous moment; represents the corresponding matrix multiplication;

[0084] The output gate uses sigmoid and tanh as activation functions to process the current input And the hidden state output of the previous moment , as follows:

[0085] (9)

[0086] in, is the output of the output gate; for Relative to The weight matrix of for Relative to The weight matrix of for The bias vector of Output the hidden state at the current moment;

[0087] Set the loss function during the training process of the LSTM neural network source prediction model L oss is:

[0088] (10)

[0089] in, represents the root mean square error; is the total number of training data; For lithium battery Actual capacity forecast value; For lithium battery The actual capacity value.

[0090] The source prediction model to be constructed can change the model structure according to the actual training results. The model theme includes an input layer, a hidden layer and an output layer connected in sequence; the number of neurons in the input layer is the sum of the number of intrinsic mode functions and residual sequences obtained after the original capacity degradation data is subjected to empirical mode decomposition; the hidden layer includes two LSTM layers connected in sequence and a dense connection layer. The two LSTM layers learn the common characteristics of lithium battery capacity degradation data, and consider the correlation between capacity degradation time series data through effective long-term dependency learning. The dense connection layer is before the output layer, in order to perform transfer learning on the source prediction model later and fine-tune the model, in order to learn the specific attributes of the target task, so that when there may be differences between the target domain data and the source data, the role of transfer learning can be played; the number of outputs in the output layer is the number of output categories of the target task.

[0091] The data obtained by data normalization and empirical mode decomposition in step S2 is input to pre-train and save the source prediction model established above for transfer learning of subsequent models. The expression of the future capacity predicted by the pre-trained source prediction model is as follows:

[0092] (6)

[0093] in, For the Future capacity predicted by cycles; is the weight matrix of the densely connected layer, and are the weight matrices corresponding to the first LSTM layer and the second LSTM layer respectively, and are the bias vectors of the densely connected layer, the first LSTM layer, and the second LSTM layer, respectively. For the current The hidden state of the LSTM layer at this moment.

[0094] The long-term dependence of the source prediction model can represent the correlation between the current capacity and the historical capacity in the long-term degradation process of lithium batteries. The capacity degradation dataset generally covers hundreds of cycles, and the degradation information between these cycles is highly correlated.

[0095] Step S4: Fine-tune the source prediction model transfer learning

[0096] The source prediction model learns the common characteristics of data in the degradation process of lithium batteries. However, sometimes the degradation process of lithium batteries is closely related to the materials, working conditions, temperature, etc. of lithium batteries. The source prediction model only learns the specific attributes contained in the source data set. When there is capacity degradation data obtained from another lithium battery working under different working conditions, temperature, etc., when the data is input into the source prediction model, due to different working conditions, different degradation stresses are caused. The source prediction model may not be suitable for the data set for future capacity prediction. Therefore, it is necessary to transfer learning of the source prediction model so that the model can adapt to the tasks of the target domain data set.

[0097] In this embodiment, a small amount of historical capacity degradation data of the target lithium battery No. B0018 for which future capacity needs to be predicted is first obtained (which can be obtained directly from the NASA laboratory) to form a target domain data set, and then the target domain data set is processed according to step S2 to obtain input data. Since the LSTM layer in the aforementioned source prediction model is to learn the inherent properties of the lithium battery degradation process, these two layers are directly copied to the optimal prediction model. For the dense layer and the output layer, the parameters are randomly initialized and added to the LSTM layer. By inputting the input data into the optimal prediction model, the parameters of the LSTM layer are fine-tuned based on the parameters of the source prediction model, and the dense connection layer and the output layer are trained from scratch, the final optimal prediction model is obtained. Since the source prediction model has learned the inherent properties of lithium battery degradation, when the target domain data is less, the generalization ability of the model can be improved by fine-tuning. The schematic diagram of the optimal prediction model obtained by transfer learning of the source prediction model in the present invention is shown in the figure. Figure 6 shown.

[0098] Step S5, predicting the future capacity of target lithium battery No. B0018.

[0099] The optimal prediction model suitable for the target domain data set is obtained by transfer learning, and then a sliding window data of the target lithium battery B0018 can be used to iteratively predict the future capacity. Specifically, in this embodiment, the historical capacity degradation data of the target lithium battery B0018 is normalized and empirical mode decomposed to construct a sliding time window data, and the time window data size is set to 10 (such as Figure 7 As shown in the figure, each window data is for the number of cycles of the lithium battery, and each time interval is 1 cycle number), the current time window data is input into the optimal prediction model to predict the future capacity of the target lithium battery in the next cycle, and the predicted future capacity is updated to the window data. A new round of sliding window data is input into the optimal prediction model to predict the future capacity of the next cycle, and the cycle is repeated. When the future capacity of the target lithium battery drops to 80% of the initial rated capacity, it is considered that the target lithium battery has reached the end of its life. At this time, the target lithium battery future capacity prediction curve is drawn and the remaining life of the target lithium battery is calculated, and the entire prediction process is completed. Figure 8 ( Figure 8 middle, From the comparison result between the predicted value of the future capacity of target lithium battery B0018 (representing the capacity at the end of the life of the lithium battery) and the actual value, it can be seen that the optimal prediction model involved in the present invention can not only better capture the overall trend of lithium battery degradation, but also capture the phenomenon of local capacity regeneration fluctuation. The model shows a higher accuracy in predicting the future capacity and remaining life of the lithium battery.

[0100] The above embodiments are preferred implementation schemes of the present invention. In addition, the present invention may also be implemented in other ways. Any obvious replacement without departing from the concept of the present technical solution is within the protection scope of the present invention.

[0101] In order to make it easier for ordinary technicians in the field to understand the improvements of the present invention over the prior art, some drawings and descriptions of the present invention have been simplified, and for the sake of clarity, some other elements are omitted in this application document. Ordinary technicians in the field should realize that these omitted elements may also constitute the content of the present invention.

Claims

1. A lithium battery capacity prediction method based on empirical mode decomposition and data fusion driven, characterized in that: include: Step S1, obtaining historical capacity degradation data of a lithium battery to form a source data set; Step S2, using empirical mode decomposition to iteratively screen the source data set, decomposing its historical capacity degradation data into a residual sequence and multiple intrinsic mode functions, wherein the high-frequency parts of capacity regeneration and capacity fluctuation are represented by the intrinsic mode function part, and the overall trend of capacity degradation is represented by the residual sequence; Step S3, constructing a long short-term memory neural network source prediction model, and pre-training the constructed source prediction model using the residual sequence and intrinsic mode function data obtained in step S2; The source prediction model includes an input layer, a hidden layer and an output layer connected in sequence; wherein the number of neurons in the input layer is the sum of the number of intrinsic mode functions and residual sequences obtained after the original capacity degradation data is subjected to empirical mode decomposition; the hidden layer includes two LSTM layers connected in sequence and a dense connection layer, the two LSTM layers learn the common characteristics of lithium battery capacity degradation data, consider the correlation between capacity degradation time series data through effective long-term dependency learning, and the dense connection layer performs transfer learning on the source prediction model to fine-tune the model; the number of outputs in the output layer is the number of output categories of the target task; The future capacity predicted by the pre-trained source prediction model is expressed as follows: （6） in, For the Future capacity predicted by cycles; is the weight matrix of the densely connected layer, and are the weight matrices corresponding to the first LSTM layer and the second LSTM layer respectively, and are the bias vectors of the densely connected layer, the first LSTM layer, and the second LSTM layer, respectively. For the current The hidden state of the LSTM layer at this moment; Step S4, fine-tuning the parameters of the pre-trained source prediction model to implement transfer learning and obtain the optimal prediction model; First, the historical capacity degradation data of the target lithium battery is obtained to form a target domain data set. Then, the target domain data set is processed according to step S2 to obtain input data. Then, two LSTM layers of the source prediction model are directly copied. Then, the dense connection layer and the output layer of the source prediction model are added to the copied LSTM layer through random parameter initialization. Finally, the parameters of the copied LSTM layer are fine-tuned based on the source prediction model using the input data. The dense connection layer and the output layer connected to the copied LSTM layer are trained from scratch to obtain the optimal prediction model. Step S5, using the optimal prediction model obtained in step S4 to predict the future capacity of the target lithium battery.

2. The lithium battery capacity prediction method based on empirical mode decomposition and data fusion drive according to claim 1 is characterized in that: In the step S1, the historical capacity degradation data of the lithium battery is obtained by monitoring the capacity of the lithium battery at different charge and discharge cycle stages.

3. The lithium battery capacity prediction method based on empirical mode decomposition and data fusion drive according to claim 2 is characterized in that: The step S2 comprises: S201, the historical capacity degradation data of lithium batteries After normalization, we get : （1） in, and They are the maximum and minimum values ​​in the historical capacity degradation data of lithium batteries; S202, using empirical mode decomposition to analyze the data Perform iterative screening, from the number of iterations To the maximum number of iterations , do the following: 1) Select Settings Data , by comparing adjacent values ​​within all fluctuations, searching All local maxima and minima in the , and then use splines to connect these local extreme values ​​to construct the upper envelope and lower envelope ; 2) Calculate the local average : （2） 3) Calculate data With the local mean difference : （3） 4) Pass inspection Whether the standard of the eigenmode function is met is determined contains the number of eigenmode functions, if If the signal meets the standard of intrinsic mode function, it is set as IMF signal. Remove this , to obtain the corresponding residual : （4） 5) Repeat steps 2) to 4) until the residual is becomes a monotonic function; if the predefined maximum number of iterations is reached , the screening cycle is terminated; after this transformation process, the intrinsic mode function contains the information of capacity regeneration and capacity fluctuation, and the data It is expressed as: （5） in, Indicates from The number of eigenmode functions decomposed from represents the jth intrinsic mode function after decomposition, Represents the residuals that form a monotonic function.

4. The lithium battery capacity prediction method based on empirical mode decomposition and data fusion drive according to claim 3 is characterized in that: In step S5, after the historical capacity degradation data of the target lithium battery is normalized and subjected to empirical mode decomposition, a sliding time window data is constructed, and the time window data size is set to 10. The current time window data is input into the optimal prediction model to predict the future capacity of the target lithium battery in the next cycle, and the predicted future capacity is updated to the window data. A new round of sliding window data is input into the optimal prediction model to predict the future capacity of the next cycle. This cycle is repeated. When the future capacity of the target lithium battery drops to 80% of the initial rated capacity, it is considered that the target lithium battery has reached the end of its life. At this time, a future capacity prediction curve of the target lithium battery is drawn and the remaining life of the target lithium battery is calculated, and the entire prediction process ends.

5. The lithium battery capacity prediction method based on empirical mode decomposition and data fusion drive according to claim 1 or 4, characterized in that: In step S1, a B0005 18650 lithium battery with a rated capacity of 2Ah is selected, and a full charge and discharge cycle test is performed under room temperature and constant current conditions. The test process is as follows: 1) Discharge the lithium battery at a constant load current of 2A. When the discharge reaches the cut-off voltage of 2.6V, stop discharging and let it stand for 2h. 2) First, charge the battery to 4.2 V with a constant current of 1.5 A, then charge it at a constant voltage of 4.2 V until the charging current is less than 0.02 A, and leave it alone for 30 min; 3) Discharge the lithium battery at a fixed load current level of 2A until the voltage reaches 2.6 V, then stop discharging and let it stand for 30 minutes; 4) Repeat steps 2) and 3) until the capacity of the lithium battery reaches the end of its life. After each discharge, record the discharge capacity of the battery. The discharge capacity is the actual rated capacity of the lithium battery.

Images