A lithium ion battery SOH prediction method based on an informer neural network

Abstract

The application discloses a lithium ion battery SOH prediction method based on an Informer neural network, and has the characteristics that the method comprises the following steps: S1, obtaining lithium battery cycle aging data to make a data set, and dividing the data set into a training set and a test set; S2, data set preprocessing, obtaining cycle aging data in the data set, and performing normalization processing; S3, constructing an initial Informer neural network; S4, improving the initial Informer neural network; S5, taking the preprocessed training set as input to train a final Informer neural network; S6, inputting the preprocessed test set into the trained final Informer neural network to obtain a prediction result, and taking mean absolute error and root mean square error as evaluation indexes. The method performs multiple correlation coefficient analysis on cycle data, determines a proper window-size, and improves the Informer neural network. A large number of experimental results show that the method can not only well extract each cycle feature with the strongest correlation with the current moment SOH, but also accurately estimate the SOH.

Description

Technical Field

[0001] This invention relates to the field of power battery management technology, specifically to a lithium-ion battery SOH prediction method based on Informer neural network. Background Technology

[0002] Lithium-ion batteries, as a clean energy source, are now widely used in electric vehicles, electronic devices, and energy storage systems. However, lithium batteries naturally age over time, and their performance may degrade or even cause safety incidents due to improper operation and other issues. Currently, the technology for assessing the state of health (SOH) of lithium batteries is not yet fully mature. On the one hand, SOH values ​​cannot be accurately measured by sensors; on the other hand, the degradation process of lithium batteries varies under different operating conditions. Therefore, from a long-term perspective, assessing the state of health (SOH) of lithium batteries is crucial for ensuring safe battery operation.

[0003] In recent years, numerous methods for estimating the state of health (SOH) of lithium-ion batteries have emerged, which can be categorized into direct measurement methods, model-based methods, and data-driven methods. Direct measurement methods require specialized personnel using specific instruments and can only be performed offline, not online. Furthermore, direct measurement methods can cause some damage to the battery and have limited accuracy. Model-based methods, on the other hand, model and simulate the internal and external characteristics of lithium-ion batteries. However, these methods have relatively weak generalization capabilities and require different models for different battery types. Moreover, relying solely on a single equivalent circuit model cannot fully reflect the internal changes within the battery, thus resulting in poor performance in practical applications.

[0004] With the explosive growth of data volume, data-driven methods have received widespread attention. These methods do not require detailed analysis of the internal chemical reactions and external characteristics of lithium batteries; they only require various parameters recorded during battery use to provide a relatively accurate estimate of the State of Harm (SOH). Currently, neural networks such as Recurrent Neural Networks (RNNs), Convolutional Neural Networks (CNNs), and Transformers are widely used for SOH estimation. Among them, Informer is an efficiency-optimized long-term series prediction model based on Transformer. However, because Informer's ProbSparse self-attention mechanism is not effective at extracting features from the large amount of serialized information in lithium batteries, this model cannot accurately estimate the SOH of lithium batteries. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of existing technologies and propose a lithium-ion battery SOH prediction method based on an Informer neural network. This method improves the Informer neural network by determining an appropriate window size through multiple correlation coefficient analysis of cyclic data. Extensive experimental results show that this method can not only effectively extract the cyclic features most strongly correlated with the current SOH, but also estimate the SOH with relatively high accuracy.

[0006] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows:

[0007] A method for predicting the state of harm (SOH) of lithium-ion batteries based on an Informer neural network includes the following steps:

[0008] S1. Obtain lithium battery cycle aging data to create a dataset, and divide it into training set and test set;

[0009] S2. Dataset preprocessing: Obtain the cyclic aging data in the dataset and perform normalization processing.

[0010] S3. Construct an initial Informer neural network, which includes an Embedding module, an Encoder module, a Decoder module, and a fully connected layer;

[0011] S4. Improved Initial Informer Neural Network

[0012] Multiple correlation coefficient analysis was performed on the preprocessed cyclic aging data to obtain the cycle number with the strongest relationship with the current cycle's SOH, and then the window-size was determined and substituted into the initial Informer neural network to obtain the final Informer neural network;

[0013] S5. Use the preprocessed training set as input to train the final Informer neural network;

[0014] S6. The preprocessed test set is input into the final Informer neural network after training to obtain the prediction results, and the mean absolute error and root mean square error are used as evaluation indicators.

[0015] Preferably, the cycle aging data includes the original voltage, current, temperature, and charge / discharge capacity.

[0016] Preferably, in step S1, the ratio of the training set to the test set is 3:1.

[0017] Preferably, the Embedding module includes scalar projection, position embedding, and global timestamp embedding; the Encoder module includes a multi-head local attention mechanism, residual and layer normalization operations; and the Decoder module includes a multi-head local attention mechanism, residual and layer normalization operations, and a multi-head self-attention layer.

[0018] As a preferred approach, the multiple correlation coefficient analysis procedure is as follows: First, the R-squared value of the multiple determination coefficient is used to assess the goodness of fit of the model. R-squared > 0.35 indicates a reliable model, R-squared > 0.5 indicates a good model fit, and R-squared > 0.7 indicates an excellent model fit. Then, the F-statistic is calculated to determine whether the significance level P <= 0.01, ensuring the reliability of the hypothesis. Finally, the regression coefficient B is calculated. * The standardized coefficient Beta is used, where the Beta value reflects the importance of the independent variable x to the dependent variable y; a larger absolute value indicates a greater influence of the independent variable on the dependent variable. This allows us to identify the cycle with the strongest influence on the current cycle's SOH (i.e., the dependent variable y) (i.e., the cycle with the largest regression coefficient among the independent variables, x). Through this analysis, we can determine the cycle number with the strongest relationship to the current cycle's SOH, thus obtaining the adjacent distance between the strongest-relationship cycle and the current cycle. Using this distance as the window size, we can further determine the window size for the improved initial Informer neural network attention mechanism.

[0019] Preferably, the training method in step S5 is as follows: where Epoch=50, learning rate lr=0.0001, batch size=32, number of encoder layers e_layers=2, number of decoder layers d_layers=1, number of attention heads n_heads=8, loss function loss=L1Loss, encoder input sequence length seq_len=96, decoder initial feature length label_len=48, sequence prediction length pred_len=12, temporal encoding freq=h, and model dimension d_model=512.

[0020] Preferably, in step S6, the method for making predictions using the final Informer neural network after training is as follows:

[0021] S6-1. Using the preprocessed test set as input, randomly select a sequence of length S=96 as the encoder input, where... , These are used as the initial loop data input and timestamp input for the encoder, respectively. For sequence dimensions, This specifies the time encoding type. Then, a portion of the sequence of length L=48 is selected from the above sequence and concatenated with a sequence of prediction length P=12 initialized to 0 as the decoder input, where... , These serve as the initial loop data input and timestamp input for the decoder, respectively. The encoder input vector is obtained through the Embedding module. and decoder input vector ;

[0022] S6-2, Input the completed embedding vector The encoded result is obtained through the Encoder module. ;

[0023] S6-3, Input the completed encoding matrix and the vector that has been embedded The decoded result is obtained through the Decoder module. ;

[0024] S6-4 Finally, the input is the decoder's output. After passing through a fully connected layer, a linear transformation and feature dimensionality reduction are performed to complete the many-to-one mapping and obtain the prediction result. The expression is as follows:

[0025]

[0026] in, This represents the decoding result of the decoder. Indicates a fully connected layer. This indicates the prediction result.

[0027] Preferably, the specific method of step S6-1 is as follows:

[0028] S6-1-1, the inputs are respectively and , here recorded as the number input sequences Feature embedding is achieved by using a one-dimensional convolutional filter with a kernel size of 3 and a stride of 1. Projected to 3D eigenvectors In the middle, the output is a scalar projection. The expression is as follows:

[0029]

[0030] in, Indicates the first Input sequences, It is the first in the vector One dimension, This represents a one-dimensional convolutional filter;

[0031] S6-1-2, Input is and , here it is recorded The input sequence is mapped to positions in the sin and cosine functions. Given a positional information on the sine or cosine function, the output is the position of the input sequence. Position embedding in each dimension The expression is as follows:

[0032]

[0033] in, yes In sequence The position in the middle, It is the first in the vector One dimension, This refers to the dimension of the model; 10000 is used to control the frequency of the sine and cosine functions, and the exponent... Control the rate of increase of frequency in a dimension;

[0034] S6-1-3, Input is and , here it is recorded Global timestamp The input timestamp is mapped to 512 dimensions through a fully connected layer, and the output is the [dimensional value]. timestamp embedding of a sequence The expression is as follows:

[0035]

[0036] in, Indicates the first Input the time information of each sequence. It is the first in the vector One dimension, This indicates different types of global timestamps, including types with minimum intervals of month, week, day, hour, and minute, i.e., the specified frequency. Indicates a fully connected layer;

[0037] S6-1-4. The input embedding consists of three independent parts: scalar projection, position embedding, and global timestamp embedding. Their calculation methods are shown in S6-1-1, S6-1-2, and S6-1-3. The output is the sum of these three parts. Input vector of a sequence The expression is as follows:

[0038]

[0039] in It is the first in the vector One dimension, Used for balancing scalar projection Location embedding and timestamp embedding Regarding the relationship between sizes, the sequence has been standardized, so we take 1.

[0040] Preferably, the specific method of step S6-2 is as follows:

[0041] S6-2-1, The input is obtained from step S6-1-4. For the nth head, the input vector is first transformed into a linear form. , , Three vectors, expressed as follows:

[0042]

[0043] in, Represents the query matrix. Represents the key matrix, Represents a value matrix. This is the weight parameter matrix obtained after passing through a linear layer. ;

[0044] The attention weights are calculated using Local-Self-Attention, with the input being the value obtained from the above formula. , , The output is the attention of the nth head. The expression is as follows:

[0045]

[0046] in, Represents the dimension of a vector. This represents a position in the input sequence aligned with that time point. Used to measure window size, i.e. Used for , The process involves truncating the input sequence; if the window exceeds the range of the input sequence, the excess portion is discarded. For activation functions;

[0047] To merge all the headers, first concatenate them column by column to get... Then multiply by the weight matrix , get output The expression is as follows:

[0048]

[0049] in, This represents multiple attention heads, each independently calculating attention weights and generating an attention head output. , For concatenation functions, It is the output weight matrix, used to perform a linear transformation on the outputs of multiple attention heads;

[0050] S6-2-2, Residual and Layer Normalization Operations, with the input obtained from step S6-1-4. and obtained from step S6-2-1 The output is The expression is as follows:

[0051]

[0052] in, Indicates input , express The result obtained after Multi-Head Local-Self-Attention , This indicates Layernorm normalization. This indicates dropout regularization;

[0053] S6-2-3, Residual and Layer Normalization Operations, with the input obtained from step S6-2-2. The output is the encoder output. The expression is as follows:

[0054]

[0055] in, This indicates Layernorm normalization. It is a feedforward neural network.

[0056] Preferably, the specific method of step S6-3 is as follows:

[0057] S6-3-1, Same as step S6-2-1, the input is... Output ;

[0058] S6-3-2, Same as step S6-2-2, the input is the result obtained in step S6-1-4. and obtained from step S6-3-1 The output is ;

[0059] S6-3-3, Input is obtained from the encoder module , The decoder module obtained The output is cross-encoded with the output of the encoder module through a multi-head self-attention layer, and the output is an attention layer. The expression is as follows:

[0060]

[0061] in, It was obtained through S6-3-1 , and It was obtained through S6-2-1 , , Represents the dimension of a vector. For activation functions;

[0062] S6-3-4, Same as step S6-2-2, the input is obtained from step S6-3-2. and obtained from step S6-3-3 The output is ;

[0063] S6-3-5, Same as step S6-2-3, the input is obtained from step S6-3-4. The output is the decoder output. .

[0064] This invention has the following characteristics and beneficial effects:

[0065] By employing the above technical solution, the cyclic data is analyzed using the multiple correlation coefficient to obtain the cycle number with the strongest correlation to the current cycle's SOH. This yields the neighbor distance between the strongest-correlated cycle and the current cycle. This distance is used as the window size to improve the attention mechanism of the Informer neural network, replacing its sparse attention mechanism, which randomly selects features, with a local attention mechanism more suitable for lithium battery SOH sequence features. For time-correlated lithium battery SOH, the improved Informer can utilize the local window size in the attention mechanism to focus on key attention components. Using the analysis results of the multiple correlation coefficient, it selects cycles with a stronger correlation to the current cycle's SOH for prediction, discarding cycles with a weaker correlation, thus saving prediction time and improving prediction accuracy. Attached Figure Description

[0066] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0067] Figure 1 This is a flowchart of a method according to an embodiment of the present invention.

[0068] Figure 2 This is a schematic diagram of the Informer neural network in an embodiment of the present invention.

[0069] Figure 3 This is a diagram showing the prediction results of the Informer neural network in an embodiment of the present invention. Detailed Implementation

[0070] The experimental environment used in this embodiment is as follows: CPU Intel(R) Xeon(R) Gold 6330 CPU @2.00GHz, GPU RTX 3090, GPU memory 24GB, Python version 3.8, CUDA version 12.0, and the deep learning framework used is PyTorch-GPU 1.8.1.

[0071] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other.

[0072] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, the terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined with "first," "second," etc., may explicitly or implicitly include one or more of that feature. In the description of this invention, unless otherwise stated, "a plurality of" means two or more.

[0073] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art will understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0074] This invention provides a method for predicting the state of harm (SOH) of lithium-ion batteries based on an Informer neural network, such as... Figure 1 As shown, it includes the following steps:

[0075] S1. Obtain lithium battery cycle aging data to create a dataset, and divide it into training set and test set;

[0076] In this embodiment, voltage data at different time points during the constant current charging phase of the battery are selected as data samples, with the ratio of training set to test set samples being 3:1.

[0077] S2. Dataset preprocessing: Obtain the cyclic aging data in the dataset and perform normalization processing.

[0078] Specifically, raw cyclic aging data such as voltage, current, temperature, and charge / discharge capacity are obtained from the dataset and normalized. Then, multiple correlation coefficient analysis is performed on the obtained cyclic data. The multiple correlation coefficient measures the performance of a variable... Other variables An indicator of the degree of linear correlation between them. By performing correlation analysis on SOH using the multiple correlation coefficient, the cycle number with the strongest relationship to the current cycle's SOH can be obtained, that is, the cycle that has the greatest impact on the SOH to be predicted.

[0079] Furthermore, the method for analyzing the multiple correlation coefficient is as follows:

[0080] First, the R-squared value (complex coefficient of determination) is used to assess the model's fit. An R-squared value > 0.35 indicates a reliable model, > 0.5 indicates a good fit, and > 0.7 indicates an excellent fit. Then, the F-statistic is calculated to determine the significance level (P < 0.01), ensuring the hypothesis is credible. Finally, the regression coefficient B is calculated. * The standardized coefficient Beta is used, where the Beta value reflects the importance of the independent variable x to the dependent variable y; a larger absolute value indicates a greater influence of the independent variable on the dependent variable. This allows us to identify the cycle with the greatest impact on the current cycle's SOH (i.e., the dependent variable y) (i.e., the cycle with the largest regression coefficient x among the independent variables). Through this analysis, we can determine the cycle number with the strongest relationship to the current cycle's SOH, thus obtaining the adjacent distance between the strongest-relationship cycle and the current cycle. Using this distance as the window size, we can further determine the window size for the improved initial Informer neural network attention mechanism.

[0081] Specifically, the multiple correlation coefficient analysis method includes the following sub-steps:

[0082] (1) Determine the dependent variable and independent variable The dependent variable is the SOH of one of the cycles, and the independent variable is the data from the remaining cycles;

[0083] (2) Establish a regression model, i.e. ;

[0084] (3) Calculate the multiple correlation coefficient R. The calculation formula is as follows:

[0085]

[0086] in, for right The result obtained from the regression is... . express The average value;

[0087] (4) Calculate the F-statistic and consult the F-distribution table to determine the significance level P. The calculation formula is as follows:

[0088]

[0089] in, The multiple correlation coefficient, The sample size, i.e., the independent variable. The number of features included This represents the number of independent variables. Consulting the F-distribution table, if the actual value of F > the table value, then the significance level P <= 0.01, indicating that the probability of the hypothesis being rejected is less than 0.01, meaning the hypothesis is credible.

[0090] (5) Calculate the regression coefficients

[0091]

[0092]

[0093]

[0094] in, The matrix of independent variables, Here is the true regression coefficient matrix, where These are the coefficients in the regression model. These are the regression coefficients obtained using the least squares method. To specify the dependent variable;

[0095] (6) Combining standardized coefficients The value is analyzed in detail, and the calculation formula is as follows:

[0096]

[0097] in, These are the calculated regression coefficients. Dependent variable The standard deviation. as independent variable The standard deviation.

[0098] Finally, the regression coefficient B is calculated. * The standardized coefficient Beta is used, where the Beta value reflects the importance of the independent variable x to the dependent variable y; a larger absolute value indicates a greater influence of the independent variable on the dependent variable. This allows us to identify the cycle with the greatest impact on the current cycle's SOH (i.e., the dependent variable y) (i.e., the cycle with the largest regression coefficient x among the independent variables). Through this analysis, we can determine the cycle number with the strongest relationship to the current cycle's SOH, thus obtaining the adjacent distance between the strongest-relationship cycle and the current cycle. Using this distance as the window size, we can further determine the window size for the improved initial Informer neural network attention mechanism.

[0099] S3. Construct the initial Informer neural network, such as Figure 2 As shown, the initial Informer neural network includes an Embedding module, an Encoder module, a Decoder module, and a fully connected layer;

[0100] The Embedding module includes scalar projection, position embedding, and global timestamp embedding; the Encoder module includes a multi-head local attention mechanism, residual and layer normalization operations; the Decoder module includes a multi-head local attention mechanism, residual and layer normalization operations, and a multi-head self-attention layer.

[0101] S4. Improved Initial Informer Neural Network

[0102] Multiple correlation coefficient analysis was performed on the preprocessed cyclic aging data to obtain the cycle number with the strongest relationship with the current cycle's SOH, and then the window-size was determined and substituted into the initial Informer neural network to obtain the final Informer neural network;

[0103] S5. Use the preprocessed training set as input to train the final Informer neural network;

[0104] During training, the following parameters were used: Epoch=50, learning rate lr=0.0001, batch size=32, encoder layers e_layers=2, decoder layers d_layers=1, attention heads n_heads=8, loss function loss=L1Loss, encoder input sequence length seq_len=96, decoder initial feature length label_len=48, sequence prediction length pred_len=12, temporal encoding freq=h, and model dimension d_model=512.

[0105] S6. The preprocessed test set is input into the final Informer neural network after training to obtain the prediction results, and the mean absolute error and root mean square error are used as evaluation indicators.

[0106] Furthermore, in step S6, the method for making predictions using the final Informer neural network after training is as follows:

[0107] S6-1. Use the preprocessed test set as input.

[0108] In this embodiment, the input to the Embedding module is a randomly selected sequence of length S=96 as the encoder input, where... , These are used as the initial loop data input and timestamp input for the encoder, respectively. For sequence dimensions, This specifies the time encoding type. Then, a portion of the sequence of length L=48 is selected from the above sequence and concatenated with a sequence of prediction length P=12 initialized to 0 as the decoder input, where... , These serve as the initial loop data input and timestamp input for the decoder, respectively. The encoder input vector is obtained through the Embedding module. and decoder input vector .

[0109] Specifically,

[0110] S6-1-1, the inputs are respectively and , here recorded as the number input sequences Feature embedding is achieved by using a one-dimensional convolutional filter with a kernel size of 3 and a stride of 1. Projected to 3D eigenvectors In the middle, the output is a scalar projection. The expression is as follows:

[0111]

[0112] in, Indicates the first Input sequences, It is the first in the vector One dimension, This represents a one-dimensional convolutional filter;

[0113] S6-1-2, Input is and , here it is recorded as The input sequence is mapped to positions in the sin and cosine functions. Given a positional information on the sine or cosine function, the output is the position of the input sequence. Position embedding in each dimension The expression is as follows:

[0114]

[0115] in, yes In sequence The position in the middle, It is the first in the vector One dimension, This refers to the dimension of the model; 10000 is used to control the frequency of the sine and cosine functions, and the exponent... Control the rate of increase of frequency in a dimension;

[0116] S6-1-3, Input is and , here it is recorded Global timestamp The input timestamp is mapped to 512 dimensions through a fully connected layer, and the output is the [dimensional value]. timestamp embedding of a sequence The expression is as follows:

[0117]

[0118] in, Indicates the first Input the time information of each sequence. It is the first in the vector One dimension, This indicates different types of global timestamps, including types with minimum intervals of month, week, day, hour, and minute, i.e., the specified frequency. Indicates a fully connected layer;

[0119] S6-1-4. The input embedding consists of three independent parts: scalar projection, position embedding, and global timestamp embedding. Their calculation methods are shown in S6-1-1, S6-1-2, and S6-1-3. The output is the sum of these three parts. Input vector of a sequence The expression is as follows:

[0120]

[0121] in It is the first in the vector One dimension, Used for balancing scalar projection Location embedding and timestamp embedding Regarding the relationship between sizes, the sequence has been standardized, so we take 1.

[0122] S6-2, Input the completed embedding vector The encoded result is obtained through the Encoder module. .

[0123] Specifically,

[0124] S6-2-1, The input is obtained from step S6-1-4. For the nth head, the input vector is first transformed into a linear form. , , Three vectors, expressed as follows:

[0125]

[0126] in, Represents the query matrix. Represents the key matrix, Represents a value matrix. This is the weight parameter matrix obtained after passing through a linear layer. ;

[0127] The attention weights are calculated using Local-Self-Attention, with the input being the value obtained from the above formula. , , The output is the attention of the nth head. The expression is as follows:

[0128]

[0129] in, Represents the dimension of a vector. This represents a position in the input sequence aligned with that time point. Used to measure window size, i.e. Used for , The process involves truncating the input sequence; if the window's range exceeds the input sequence's range, the excess portion is discarded. For activation functions;

[0130] To merge all the headers, first concatenate them column by column to get... Then multiply by the weight matrix , get output The expression is as follows:

[0131]

[0132] in, This represents multiple attention heads, each independently calculating attention weights and generating an attention head output. , For concatenation functions, It is the output weight matrix, used to perform a linear transformation on the outputs of multiple attention heads;

[0133] S6-2-2, Residual and Layer Normalization Operations, with the input obtained from step S6-1-4. and obtained from step S6-2-1 The output is The expression is as follows:

[0134]

[0135] in, Indicates input , express The result obtained after Multi-Head Local-Self-Attention , This indicates Layernorm normalization. This indicates dropout regularization;

[0136] S6-2-3, Residual and Layer Normalization Operations, with the input obtained from step S6-2-2. The output is the encoder output. The expression is as follows:

[0137]

[0138] in, This indicates Layernorm normalization. It is a feedforward neural network.

[0139] S6-3, Input the completed encoding matrix and the vector that has been embedded The decoded result is obtained through the Decoder module. ;

[0140] S6-3-1, Same as step S6-2-1, the input is... Output ;

[0141] S6-3-2, Same as step S6-2-2, the input is the result obtained in step S6-1-4. and obtained from step S6-3-1 The output is ;

[0142] S6-3-3, Input is obtained from the encoder module , The decoder module obtained The output is cross-encoded with the output of the encoder module through a multi-head self-attention layer, and the output is an attention layer. The expression is as follows:

[0143]

[0144] in, It was obtained through S6-3-1 , and It was obtained through S6-2-1 , , Represents the dimension of a vector. For activation functions;

[0145] S6-3-4, Same as step S6-2-2, the input is obtained from step S6-3-2. and obtained from step S6-3-3 The output is ;

[0146] S6-3-5, Same as step S6-2-3, the input is obtained from step S6-3-4. The output is the decoder output. .

[0147] S6-4 Finally, the input is the decoder's output. After passing through a fully connected layer, a linear transformation and feature dimensionality reduction are performed to complete the many-to-one mapping and obtain the prediction result. The expression is as follows:

[0148]

[0149] in, This represents the decoding result of the decoder. Indicates a fully connected layer. Indicates the prediction result;

[0150] S6-5. Conduct experiments on public datasets.

[0151] The test data was input into the network model trained with S5, and a comparative experiment was conducted with CNN and LSTM network models, such as... Figure 3 As shown in the table below:

[0152]

[0153] The evaluation metrics used are mean absolute error (MAE) and root mean square error (RMSE). For the proposed method, the MAE is 1.78% and the RMSE is 2.38%. Compared with CNN and LSTM networks, the MAE and RMSE are improved by 0.71%, 1.88%, 1.64%, and 2.25%, respectively, showing that the proposed method has a significant improvement in prediction performance compared with commonly used data-driven methods. Therefore, it can be proven that the proposed lithium-ion battery SOH prediction method based on the improved Informer neural network with multiple correlation coefficient has a good model fit.

[0154] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings, but the present invention is not limited to the described embodiments. For those skilled in the art, various changes, modifications, substitutions, and variations can be made to these embodiments, including components, without departing from the principles and spirit of the present invention, and these variations still fall within the protection scope of the present invention.

Claims

1. A method for predicting the state of harm (SOH) of lithium-ion batteries based on an Informer neural network, characterized in that, Includes the following steps: S1. Obtain lithium battery cycle aging data to create a dataset, and divide it into training set and test set; S2. Dataset preprocessing: Obtain the cyclic aging data in the dataset and perform normalization processing. S3. Construct an initial Informer neural network, which includes an Embedding module, an Encoder module, a Decoder module, and a fully connected layer; S4. Improve the initial Informer neural network; Multiple correlation coefficient analysis was performed on the preprocessed cyclic aging data to obtain the cycle number with the strongest relationship with the current cycle's SOH, and then the window-size was determined and substituted into the initial Informer neural network to obtain the final Informer neural network; The method for analyzing the multiple correlation coefficient is as follows: First, the R-squared value of the multiple determination coefficient is used to assess the goodness of fit of the model. R-squared > 0.35 indicates a reliable model, R-squared > 0.5 indicates a good fit, and R-squared > 0.7 indicates an excellent fit. Then, the F-statistic is calculated to determine whether the significance level P <= 0.01, ensuring the hypothesis is credible. Finally, the regression coefficient B is calculated. * The standardized coefficient Beta is used to determine the importance of the independent variable x to the dependent variable y. The larger the absolute value of Beta, the greater the influence of the independent variable on the dependent variable. This helps to identify the cycle that has the greatest impact on the current cycle's SOH, and then determine the cycle number that has the strongest relationship with the current cycle's SOH. This gives us the distance between the cycle with the strongest relationship and the current cycle. We use this distance as the window size to further determine the window size of the improved initial Informer neural network attention mechanism. S5. The preprocessed training set is used as input to train the final Informer neural network. S6. The preprocessed test set is input into the final Informer neural network after training to obtain the prediction results, and the mean absolute error and root mean square error are used as evaluation indicators.

2. The lithium-ion battery SOH prediction method based on Informer neural network according to claim 1, characterized in that, The cycle aging data includes the original voltage, current, temperature, and charge / discharge capacity.

3. The lithium-ion battery SOH prediction method based on Informer neural network according to claim 1, characterized in that, In step S1, the ratio of the training set to the test set is 3:

1.

4. The lithium-ion battery SOH prediction method based on Informer neural network according to claim 1, characterized in that, The Embedding module includes scalar projection, position embedding, and global timestamp embedding; the Encoder module includes a multi-head local attention mechanism, residual and layer normalization operations; the Decoder module includes a multi-head local attention mechanism, residual and layer normalization operations, and a multi-head self-attention layer.

5. The lithium-ion battery SOH prediction method based on Informer neural network according to claim 1, characterized in that, The training method in step S5 is as follows: where Epoch=50, learning rate lr=0.0001, batch size=32, encoder layers e_layers=2, decoder layers d_layers=1, attention heads n_heads=8, loss function loss=L1Loss, encoder input sequence length seq_len=96, decoder initial feature length label_len=48, sequence prediction length pred_len=12, temporal encoding freq=h, and model dimension d_model=512.

6. The lithium-ion battery SOH prediction method based on Informer neural network according to claim 4, characterized in that, In step S6, the method for making predictions using the final Informer neural network after training is as follows: S6-1. Using the preprocessed test set as input, obtain the encoder input vector through the Embedding module. and decoder input vector ; S6-2. Input the encoder vector and obtain the encoded matrix through the Encoder module; S6-3. Input the encoding matrix output by the Encoder module and the decoder input vector, and obtain the decoded result through the Decoder module; S6-4 Finally, the decoding matrix output by the input Decoder module is passed through a fully connected layer for linear transformation and feature dimensionality reduction to complete the many-to-one mapping.

7. The lithium-ion battery SOH prediction method based on Informer neural network according to claim 6, characterized in that, The specific method for step S6-1 is as follows: S6-1-1, Input the initial loop data of the encoder in the test set. Initial loop data input for the decoder , here recorded as the number input sequences By using a one-dimensional convolutional filter with a kernel size of 3 and a stride of 1, feature embedding is completed. Projected to 3D eigenvectors In the middle, the output is a scalar projection. The expression is as follows: in, Indicates the first Input sequences, It is the first in the vector One dimension, This represents a one-dimensional convolutional filter; S6-1-2, Input is and , here it is recorded The input sequence is mapped to positions in the sin and cosine functions. Given a positional information on the sine or cosine function, the output is the position of the input sequence. Position embedding in each dimension The expression is as follows: in, yes In sequence The position in the middle, It is the first in the vector One dimension, This refers to the dimension of the model; 10000 is used to control the frequency of the sine and cosine functions, and the exponent... Control the rate of increase of frequency in a dimension; S6-1-3, Input is and , here it is recorded Global timestamp The input timestamp is mapped to 512 dimensions through a fully connected layer, and the output is the [dimensional value]. timestamp embedding of a sequence The expression is as follows: in, Indicates the first Input the time information of each sequence. It is the first in the vector One dimension, This indicates different types of global timestamps, including types with minimum intervals of month, week, day, hour, and minute, i.e., the specified frequency. Indicates a fully connected layer; S6-1-4. The input embedding consists of three independent parts: scalar projection, position embedding, and global timestamp embedding. The output is the sum of these three parts. Input vector of a sequence .

8. The lithium-ion battery SOH prediction method based on Informer neural network according to claim 7, characterized in that, The specific method for step S6-2 is as follows: S6-2-1, The input is obtained from step S6-1-4. For the nth head, the input vector is first transformed into a linear form. , , Three vectors, expressed as follows: in, Represents the query matrix. Represents the key matrix, Represents a value matrix, This is the weight parameter matrix obtained after passing through a linear layer. ; The attention weights are calculated using Local-Self-Attention, with the input being the value obtained from the above formula. , , The output is the attention of the nth head. The expression is as follows: in, Represents the dimension of a vector. This represents a position in the input sequence aligned with that time point. Used to measure window size, i.e. Used for , The process involves truncating the input sequence; if the window exceeds the range of the input sequence, the excess portion is discarded. For activation functions; To merge all the headers, first concatenate them column by column to get... Then multiply by the weight matrix , get output The expression is as follows: in, This represents multiple attention heads, each independently calculating attention weights and generating an attention head output. , For concatenation functions, It is the output weight matrix, used to perform a linear transformation on the outputs of multiple attention heads; S6-2-2, Residual and Layer Normalization Operations, with the input obtained from step S6-1-4. and obtained from step S6-2-1 The output is The expression is as follows: in, Indicates input , express The result obtained after Multi-Head Local-Self-Attention , This indicates Layernorm normalization. This indicates dropout regularization; S6-2-3, Residual and Layer Normalization Operations, with the input obtained from step S6-2-2. The output is the encoder output. The expression is as follows: in, This indicates Layernorm normalization. It is a feedforward neural network.

9. The lithium-ion battery SOH prediction method based on Informer neural network according to claim 8, characterized in that, The specific method for step S6-3 is as follows: S6-3-1, Same as step S6-2-1, the input is... Output ; S6-3-2, Same as step S6-2-2, the input is the result obtained in step S6-1-4. and obtained from step S6-3-1 The output is ; S6-3-3, Input is obtained from the encoder module , The decoder module obtained The output is cross-encoded with the output of the encoder module through a multi-head self-attention layer, and the output is an attention layer. The expression is as follows: in, It was obtained through S6-3-1 , and It was obtained through S6-2-1 , , Represents the dimension of a vector. For activation functions; S6-3-4, Same as step S6-2-2, the input is obtained from step S6-3-2. and obtained from step S6-3-3 The output is ; S6-3-5, Same as step S6-2-3, the input is obtained from step S6-3-4. The output is the decoder output. .

Images