Bayesian regression and dynamic state estimation fusion-based battery life prediction method

Through Bayesian regression and dynamic state estimation combined with random forest methods, the problem of inaccurate prediction of lithium battery life in the prior art is solved, accurate prediction and dynamic update of lithium battery life is achieved, and the accuracy of prediction is improved.

CN119064813APending Publication Date: 2024-12-03HEILONGJIANG ELECTRIC POWER SCIENCE RESEARCH INSTITUTE +1
View PDF 0 Cites 2 Cited by

Patent Information

Application Number
CN202411121740.0
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-15
Publication Date
2024-12-03

AI Technical Summary

Technical Problem

The existing lithium battery life prediction methods fail to accurately consider changes in charge and discharge parameters, resulting in inaccurate prediction results.

Method used

Using the battery life prediction method based on Bayesian regression and dynamic state estimation, the Bayesian network topology is determined by obtaining the charging and discharging data of lithium batteries, and the battery life prediction model is obtained by using the maximum likelihood estimation method, and combined with the random forest pairs The feature variables are sorted and monitored for importance, and the prediction model is dynamically updated.

Benefits of technology

Accurate prediction of lithium battery life is achieved, and the prediction accuracy is improved by dynamically updating the model to deal with changes in characteristic variables.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure BDA0004995293460000021
    Figure BDA0004995293460000021
  • Figure BDA0004995293460000033
    Figure BDA0004995293460000033
  • Figure BDA0004995293460000034
    Figure BDA0004995293460000034
Patent Text Reader

Abstract

The invention discloses a battery life prediction method based on Bayesian regression and dynamic state estimation fusion, relates to the technical field of lithium batteries, and aims to solve the problem of inaccurate lithium battery life prediction result in the prior art. According to the method, the feature variables are subjected to importance ranking through the random forest, the feature variables with the importance ranking larger than the threshold value are monitored, when the feature variables with the importance ranking larger than the threshold value are changed, the battery life prediction model is dynamically updated, and then accurate prediction of the battery life is achieved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of lithium batteries, and in particular to a battery life prediction method based on the fusion of Bayesian regression and dynamic state estimation. Background Art

[0002] Lithium batteries have the characteristics of high output voltage, stable discharge, fast charging speed and long life, and are widely used in various fields. With the use of lithium batteries, the increase in the number of charging cycles and complex usage, the lithium ion activity of lithium batteries will also decrease, and the power will also decrease accordingly, thus shortening the service life of lithium batteries. This will directly lead to reduced working efficiency of equipment loaded with lithium batteries, or even failure to work, and serious safety hazards. Therefore, the prediction of the remaining life of lithium batteries helps to improve the safety of battery-powered equipment.

[0003] The existing lithium battery life prediction method only uses the original lithium battery charge and discharge data to build a prediction model, and uses the prediction model for prediction. It does not consider the impact of changes in lithium battery charge and discharge parameters on the model, which leads to inaccurate lithium battery life prediction results. Summary of the invention

[0004] The purpose of the present invention is to propose a battery life prediction method based on the fusion of Bayesian regression and dynamic state estimation to address the problem of inaccurate lithium battery life prediction results in the prior art.

[0005] The technical solution adopted by the present invention to solve the above technical problems is:

[0006] The battery life prediction method based on Bayesian regression and dynamic state estimation fusion includes the following steps:

[0007] Step 1: Obtain the charge and discharge data and corresponding state of charge of the same model of lithium battery;

[0008] Step 2: Based on the lithium battery charging and discharging data, determine the Bayesian network topology structure, and use the lithium battery charging and discharging data to obtain the Bayesian network parameters using the maximum likelihood estimation method, and then obtain the battery life prediction model, and use the battery life prediction model as the current prediction model;

[0009] Step 3: Use the charge and discharge data as input feature variables and the state of charge as the target variable, input them into the random forest model, and sort the input feature variables by importance;

[0010] Step 4: Select feature variables with a ranking greater than the threshold according to the importance ranking;

[0011] Step 5: Use the current prediction model to predict the state of charge of the lithium battery, and then obtain the remaining life of the lithium battery, and monitor the selected characteristic variables. When the selected characteristic variables all change, and the change amount of each characteristic variable exceeds the respective threshold, the charging and discharging data of the lithium battery at this time is obtained and used as a new training sample;

[0012] Step 6: Based on the battery life prediction model and combined with the new training samples, the Bayesian estimation method is used to correct the Bayesian network parameters to obtain an updated battery life prediction model. The updated battery life prediction model is used as the current prediction model, and step 5 is repeated to complete the battery life prediction.

[0013] Furthermore, the charge and discharge data include the number of charge and discharge cycles, charge and discharge rate, temperature, discharge depth and charging frequency.

[0014] Furthermore, the specific steps of ranking the importance of the input feature variables in step 3 are:

[0015] The input feature variables and target variables are input into the random forest model, and the input feature variables are ranked according to the variable importance measurement method. The formula is as follows:

[0016]

[0017] Among them, VIM(x i ) is the feature x i The variable importance of g is the number of decision trees in the random forest, E t and E t (x i ) are random permutations x i Data error rate before and after.

[0018] Furthermore, in step 1, the state of charge is an actual value or is obtained through a neural network, and the steps for obtaining the state of charge through the neural network are:

[0019] The voltage, current and temperature data of the lithium battery are obtained, and the voltage, current and temperature data are input into the trained neural network to obtain the output state of charge.

[0020] Furthermore, the trained neural network is a CNN2d-LSTM-Attention model, and the CNN2d-LSTM-Attention model includes an input layer, a convolutional layer, an LSTM-Attention layer, and a fully connected layer;

[0021] The input layer is used to set the time step to 24 and the feature dimension to 12. At this time, the dimension of the data is (None, 24, 12). Then, a dimension representing the depth is added to the data. At this time, the data dimension is (None, 24, 12, 1).

[0022] The convolutional layer is used to perform convolution, batch normalization and pooling on the data output by the input layer in sequence.

[0023] Among them, the convolution uses two-dimensional Same convolution, the size of the convolution kernel is 3×3, the depth of the convolution kernel is 8, the moving step is 1, and the dimension of the data after the convolution layer is (None, 24, 12, 8). Then L2 regularization is used to constrain the weight of the convolution kernel, and the sequence extracted by the convolution kernel is batch normalized and Relu activated before being sent to the pooling layer.

[0024] The pooling adopts average pooling, and the average value in each pooling window is used as the output of the pooling layer. The size of the pooling window is 3×3, the pooling step is 3, and the dimension of the data after the pooling layer is (None, 8, 4, 8). Finally, a convolution layer with a convolution kernel size of 1×1, a moving step of 1, and a convolution kernel depth of 1 is used to adjust the number of channels. At this time, the data dimension is (None, 8, 4, 1);

[0025] The LSTM-Attention layer includes: a first LSTM layer, a multi-head attention mechanism and a second LSTM. The LSTM-Attention layer specifically performs the following steps:

[0026] The output of the convolutional layer is put into the first LSTM layer with 64 neurons to obtain the output data of each time step, that is, the dimension is (None,8,64). After that, the number of heads in the multi-head attention mechanism is set to 4, and three sets of weight matrices W with dimensions (8,8) are randomly generated according to the time step. q , W k , W v , then perform matrix multiplication on the output data of each time step, perform linear transformation to obtain three corresponding attention mechanism parameters Q, K, V, the dimensions of Q, K, V are (8,64), then divide the parameters Q, K, V into four equal parts according to the second component of the dimension, that is, 64, to obtain four head parameters with dimensions (8,16), then perform scaled dot product calculation on each head parameter to obtain the attention score S, and pass S through the softmax function to obtain the attention weight, whose dimension is (8,8), then multiply the attention weight by the matrix V to obtain the weighted sum in the head, repeat the above steps, merge the results of the weighted sums in each head, and use the merged result as the output of the final multi-head self-attention layer, the output dimension is (None,8,64);

[0027] The output of the final multi-head self-attention layer is input into the second LSTM layer, and the output of the final time step is returned, and the output dimension is (None, 64); the fully connected layer includes two Dense layers, and the data dimensions of the outputs of the two Dense layers are (None, 32) and (None, 1) respectively.

[0028] Furthermore, the specific steps of obtaining the Bayesian network parameters by using the maximum likelihood estimation method in step 2 and then obtaining the battery life prediction model are as follows:

[0029] Assumptions is a vector composed of all parameters, n is the number of nodes, q i is π(X i ), r i For node X i The number of values ​​of θ ijk =P(X i = k / π(X i )=j) is when X i The parent node value is the jth value, X i The probability of taking the kth value, D i , i=1,2,...,m is the sample data, then the vector The log-likelihood function of is:

[0030]

[0031] When θ ijk The log-likelihood function reaches its maximum value when the following values ​​are taken:

[0032]

[0033] Among them, m ijk Is the charge and discharge data that satisfies X i = k, π(X i ) = the number of samples of j; if the charge and discharge data source Then the parameters are assumed to be uniformly distributed; the Bayesian network parameters are determined by formula (2), thereby obtaining the battery life prediction model.

[0034] Furthermore, the specific steps of using the Bayesian estimation method to correct the Bayesian network parameters in step 6 are:

[0035] Assumptions Reason The word vector composed of ijk For the prior knowledge that satisfies X i = k and π(X i ) = the number of samples of j, Dirichlet distribution According to the Bayesian formula:

[0036]

[0037] in, For vector The prior probability distribution of For vector The posterior probability distribution of Dirichlet distribution therefore,

[0038]

[0039] Use the above formula to modify the Bayesian network parameters, where m ijk For new samples that satisfy X i = k and π(X i ) = the number of samples of j.

[0040] The beneficial effects of the present invention are:

[0041] The present application uses a Bayesian network to obtain a battery life prediction model to predict the charge state. The present application also uses a random forest to rank the importance of feature variables and monitors those with an importance ranking greater than a threshold. When a feature variable with an importance ranking greater than the threshold changes, the battery life prediction model is dynamically updated, thereby achieving accurate prediction of the battery life. DETAILED DESCRIPTION

[0042] It should be particularly noted that, in the absence of conflict, the various embodiments disclosed in this application can be combined with each other.

[0043] Specific implementation method 1: The battery life prediction method based on Bayesian regression and dynamic state estimation fusion described in this implementation method includes the following steps:

[0044] Step 1: Obtain the charge and discharge data and corresponding state of charge of the same model of lithium battery;

[0045] Step 2: Based on the lithium battery charging and discharging data, determine the Bayesian network topology structure, and use the lithium battery charging and discharging data to obtain the Bayesian network parameters using the maximum likelihood estimation method, and then obtain the battery life prediction model, and use the battery life prediction model as the current prediction model;

[0046] Step 3: Use the charge and discharge data as input feature variables and the state of charge as the target variable, input them into the random forest model, and sort the input feature variables by importance;

[0047] Step 4: Select feature variables with a ranking greater than the threshold according to the importance ranking;

[0048] Step 5: Use the current prediction model to predict the state of charge of the lithium battery, and then obtain the remaining life of the lithium battery, and monitor the selected characteristic variables. When the selected characteristic variables all change, and the change amount of each characteristic variable exceeds the respective threshold, the charging and discharging data of the lithium battery at this time is obtained and used as a new training sample;

[0049] Step 6: Based on the battery life prediction model and combined with the new training samples, the Bayesian estimation method is used to correct the Bayesian network parameters to obtain an updated battery life prediction model. The updated battery life prediction model is used as the current prediction model, and step 5 is repeated to complete the battery life prediction.

[0050] Assuming that both charging and discharging data have an impact and the impacts are independent of each other, the Bayesian network topology is determined; after determining the Bayesian network topology, the data obtained in the previous step is used to obtain the Bayesian network parameters using the maximum likelihood estimation method, thereby obtaining a lithium battery life prediction model.

[0051] Specific implementation method 2: This implementation method is a further explanation of specific implementation method 1. The difference between this implementation method and specific implementation method 1 is that the charge and discharge data includes the number of charge and discharge cycles, charge and discharge rate, temperature, discharge depth and charging frequency.

[0052] Specific implementation method three: This implementation method is a further description of specific implementation method two. The difference between this implementation method and specific implementation method two is that the specific steps of ranking the importance of input feature variables in step three are:

[0053] The input feature variables and target variables are input into the random forest model, and the input feature variables are ranked according to the variable importance measurement method. The formula is as follows:

[0054]

[0055] Among them, VIM(x i ) is the feature x i The variable importance of g is the number of decision trees in the random forest, E t and E t (x i ) are random permutations x i Data error rate before and after.

[0056] Specific implementation method 4: This implementation method is a further explanation of specific implementation method 3. The difference between this implementation method and specific implementation method 3 is that the state of charge in step 1 is an actual value or obtained through a neural network. The steps for obtaining the state of charge through a neural network are:

[0057] The voltage, current and temperature data of the lithium battery are obtained, and the voltage, current and temperature data are input into the trained neural network to obtain the output state of charge.

[0058] Specific implementation mode five: This implementation mode is a further description of specific implementation mode four. The difference between this implementation mode and specific implementation mode four is that the trained neural network is a CNN2d-LSTM-Attention model, and the CNN2d-LSTM-Attention model includes an input layer, a convolution layer, an LSTM-Attention layer, and a fully connected layer;

[0059] The input layer is used to set the time step to 24 and the feature dimension to 12. At this time, the dimension of the data is (None, 24, 12). Then, a dimension representing the depth is added to the data. At this time, the data dimension is (None, 24, 12, 1).

[0060] The convolutional layer is used to perform convolution, batch normalization and pooling on the data output by the input layer in sequence.

[0061] Among them, the convolution uses two-dimensional Same convolution, the size of the convolution kernel is 3×3, the depth of the convolution kernel is 8, the moving step is 1, and the dimension of the data after the convolution layer is (None, 24, 12, 8). Then L2 regularization is used to constrain the weight of the convolution kernel, and the sequence extracted by the convolution kernel is batch normalized and Relu activated before being sent to the pooling layer.

[0062] The pooling adopts average pooling, and the average value in each pooling window is used as the output of the pooling layer. The size of the pooling window is 3×3, the pooling step is 3, and the dimension of the data after the pooling layer is (None, 8, 4, 8). Finally, a convolution layer with a convolution kernel size of 1×1, a moving step of 1, and a convolution kernel depth of 1 is used to adjust the number of channels. At this time, the data dimension is (None, 8, 4, 1);

[0063] The LSTM-Attention layer includes: a first LSTM layer, a multi-head attention mechanism and a second LSTM. The LSTM-Attention layer specifically performs the following steps:

[0064] The output of the convolutional layer is put into the first LSTM layer with 64 neurons to obtain the output data of each time step, that is, the dimension is (None,8,64). After that, the number of heads in the multi-head attention mechanism is set to 4, and three sets of weight matrices W with dimensions (8,8) are randomly generated according to the time step. q , W k , W v, then perform matrix multiplication on the output data of each time step, perform linear transformation to obtain three corresponding attention mechanism parameters Q, K, V, the dimensions of Q, K, V are (8,64), then divide the parameters Q, K, V into four equal parts according to the second component of the dimension, that is, 64, to obtain four head parameters with dimensions (8,16), then perform scaled dot product calculation on each head parameter to obtain the attention score S, and pass S through the softmax function to obtain the attention weight, whose dimension is (8,8), then multiply the attention weight by the matrix V to obtain the weighted sum in the head, repeat the above steps, merge the results of the weighted sums in each head, and use the merged result as the output of the final multi-head self-attention layer, the output dimension is (None,8,64);

[0065] The output of the final multi-head self-attention layer is input into the second LSTM layer, and the output of the final time step is returned, and the output dimension is (None, 64); the fully connected layer includes two Dense layers, and the data dimensions of the outputs of the two Dense layers are (None, 32) and (None, 1) respectively.

[0066] Specific implementation method 6: This implementation method is a further explanation of specific implementation method 5. The difference between this implementation method and specific implementation method 5 is that in step 2, the maximum likelihood estimation method is used to obtain the Bayesian network parameters, and then the specific steps of obtaining the battery life prediction model are as follows:

[0067] Assumptions is a vector composed of all parameters, n is the number of nodes, q i is π(X i ), r i For node X i The number of values ​​of θ ijk =P(X i = k / π(X i )=j) is when X i The parent node value is the jth value, X i The probability of taking the kth value, D i , i=1,2,...,m is the sample data, then the vector The log-likelihood function of is:

[0068]

[0069] When θ ijk The log-likelihood function reaches its maximum value when the following values ​​are taken:

[0070]

[0071] Among them, m ijk Is the charge and discharge data that satisfies Xi = k, π(X i ) = the number of samples of j; if the charge and discharge data source Then the parameters are assumed to be uniformly distributed; the Bayesian network parameters are determined by formula (2), thereby obtaining the battery life prediction model.

[0072] Specific implementation method 7: This implementation method is a further explanation of specific implementation method 6. The difference between this implementation method and specific implementation method 6 is that the specific steps of using the Bayesian estimation method to correct the Bayesian network parameters in step 6 are:

[0073] Assumptions Reason The word vector composed of ijk For the prior knowledge that satisfies X i = k and π(X i ) = the number of samples of j, Dirichlet distribution According to the Bayesian formula:

[0074]

[0075] in, For vector The prior probability distribution of For vector The posterior probability distribution of Dirichlet distribution therefore,

[0076]

[0077] Use the above formula to modify the Bayesian network parameters, where m ijk For new samples that satisfy X i = k and π(X i ) = the number of samples of j.

[0078] It should be noted that the specific implementation is only an explanation and description of the technical solution of the present invention, and cannot be used to limit the scope of protection of the rights. Any partial changes made according to the claims and description of the present invention should still fall within the scope of protection of the present invention.

Claims

1. Battery life prediction method based on Bayesian regression and dynamic state estimation fusion, Features The following steps are involved: Step 1: Obtain the charge and discharge data and corresponding state of charge of the same model of lithium battery; Step 2: Based on the lithium battery charging and discharging data, determine the Bayesian network topology structure, and use the lithium battery charging and discharging data to obtain the Bayesian network parameters using the maximum likelihood estimation method, and then obtain the battery life prediction model, and use the battery life prediction model as the current prediction model; Step 3: Use the charge and discharge data as input feature variables and the state of charge as the target variable, input them into the random forest model, and sort the input feature variables by importance; Step 4: Select feature variables with a ranking greater than the threshold according to the importance ranking; Step 5: Use the current prediction model to predict the state of charge of the lithium battery, and then obtain the remaining life of the lithium battery, and monitor the selected characteristic variables. When the selected characteristic variables all change, and the change amount of each characteristic variable exceeds the respective threshold, the charging and discharging data of the lithium battery at this time is obtained and used as a new training sample; Step 6: Based on the battery life prediction model and combined with the new training samples, the Bayesian estimation method is used to correct the Bayesian network parameters to obtain an updated battery life prediction model. The updated battery life prediction model is used as the current prediction model, and step 5 is repeated to complete the battery life prediction.

2. The battery life prediction method based on Bayesian regression and dynamic state estimation fusion according to claim 1, Features The charge and discharge data include the number of charge and discharge cycles, charge and discharge rate, temperature, discharge depth and charging frequency.

3. The battery life prediction method based on Bayesian regression and dynamic state estimation fusion according to claim 1, Features The specific steps of ranking the importance of input feature variables in step 3 are: The input feature variables and target variables are input into the random forest model, and the input feature variables are ranked according to the variable importance measurement method. The formula is as follows: Among them, VIM(x i ) is the feature x i The variable importance of g is the number of decision trees in the random forest, E t and E t (x i ) are random permutations x i Data error rate before and after.

4. The battery life prediction method based on Bayesian regression and dynamic state estimation fusion according to claim 1, Features In step 1, the state of charge is an actual value or is obtained through a neural network. The steps for obtaining the state of charge through the neural network are: The voltage, current and temperature data of the lithium battery are obtained, and the voltage, current and temperature data are input into the trained neural network to obtain the output state of charge.

5. The battery life prediction method based on Bayesian regression and dynamic state estimation fusion according to claim 4, Features The trained neural network is a CNN2d-LSTM-Attention model, which includes an input layer, a convolutional layer, an LSTM-Attention layer, and a fully connected layer; The input layer is used to set the time step to 24 and the feature dimension to 12. At this time, the dimension of the data is (None, 24, 12). Then, a dimension representing the depth is added to the data. At this time, the data dimension is (None, 24, 12, 1). The convolutional layer is used to perform convolution, batch normalization and pooling on the data output by the input layer in sequence. Among them, the convolution uses two-dimensional Same convolution, the size of the convolution kernel is 3×3, the depth of the convolution kernel is 8, the moving step is 1, and the dimension of the data after the convolution layer is (None, 24, 12, 8). Then L2 regularization is used to constrain the weight of the convolution kernel, and the sequence extracted by the convolution kernel is batch normalized and Relu activated before being sent to the pooling layer. The pooling adopts average pooling, and the average value in each pooling window is used as the output of the pooling layer. The size of the pooling window is 3×3, the pooling step is 3, and the dimension of the data after the pooling layer is (None, 8, 4, 8). Finally, a convolution layer with a convolution kernel size of 1×1, a moving step of 1, and a convolution kernel depth of 1 is used to adjust the number of channels. At this time, the data dimension is (None, 8, 4, 1); The LSTM-Attention layer includes: a first LSTM layer, a multi-head attention mechanism and a second LSTM. The LSTM-Attention layer specifically performs the following steps: The output of the convolutional layer is put into the first LSTM layer with 64 neurons to obtain the output data of each time step, that is, the dimension is (None,8,64). After that, the number of heads in the multi-head attention mechanism is set to 4, and three sets of weight matrices W with dimensions (8,8) are randomly generated according to the time step. q , W k , W v , then perform matrix multiplication on the output data of each time step, perform linear transformation to obtain three corresponding attention mechanism parameters Q, K, V, the dimensions of Q, K, V are (8,64), then divide the parameters Q, K, V into four equal parts according to the second component of the dimension, that is, 64, to obtain four head parameters with dimensions (8,16), then perform scaled dot product calculation on each head parameter to obtain the attention score S, and pass S through the softmax function to obtain the attention weight, whose dimension is (8,8), then multiply the attention weight by the matrix V to obtain the weighted sum in the head, repeat the above steps, merge the results of the weighted sums in each head, and use the merged result as the output of the final multi-head self-attention layer, the output dimension is (None,8,64); The output of the final multi-head self-attention layer is input into the second LSTM layer, and the output of the final time step is returned, and the output dimension is (None, 64); the fully connected layer includes two Dense layers, and the data dimensions of the outputs of the two Dense layers are (None, 32) and (None, 1) respectively.

6. The battery life prediction method based on Bayesian regression and dynamic state estimation fusion according to claim 1, Features In step 2, the maximum likelihood estimation method is used to obtain the Bayesian network parameters, and then the specific steps of obtaining the battery life prediction model are as follows: Assumptions is a vector composed of all parameters, n is the number of nodes, q i is π(X i ), r i For node X i The number of values ​​of θ ijk =P(X i = k / π(X i )=j) is when X i The parent node value is the jth value, X i The probability of taking the kth value, D i , i=1,2,...,m is the sample data, then the vector The log-likelihood function of is: When θ ijk The log-likelihood function reaches its maximum value when the following values ​​are taken: Among them, m ijk Is the charge and discharge data that satisfies X i = k, π(X i ) = the number of samples of j; if the charge and discharge data source Then the parameters are assumed to be uniformly distributed; the Bayesian network parameters are determined by formula (2), thereby obtaining the battery life prediction model.

7. The battery life prediction method based on Bayesian regression and dynamic state estimation fusion according to claim 6, Features The specific steps of using the Bayesian estimation method to correct the Bayesian network parameters in step 6 are: Assumptions By θ ij1 ,θ ij2 ,...,θ ijri The word vector composed of ijk For the prior knowledge that satisfies X i = k and π(X i ) = the number of samples of j, is the Dirichlet distribution D[α ij1 ,α ij2 ,...,α ijri ], according to the Bayesian formula: in, For vector The prior probability distribution of For vector The posterior probability distribution of Follow Dirichlet distribution D[m ij1 +α ij1 ,m ij2 +α ij2 ,...,m ijri +α ijri ],therefore, Use the above formula to modify the Bayesian network parameters, where m ijk For new samples that satisfy X i = k and π(X i ) = the number of samples of j.

Citation Information

Cited By

  • Battery thermal runaway early warning system and method

    CN120184418A

  • Fused multivariate bayesian anchored attention wheelchair lithium battery rul prediction method

    CN123009662A