Lithium ion battery performance evaluation-oriented time sequence modeling method

Abstract

The invention discloses a time sequence modeling method for lithium ion battery performance evaluation, and aims to solve the technical challenges of a traditional method in the aspects of multi-scale feature extraction and fusion. The method specifically comprises the steps of obtaining battery capacity data and performing preprocessing; constructing an MTH3 model by combining a time convolutional network and a multi-head attention mechanism on the basis of a Mama model, and performing multi-time scale feature extraction by taking the preprocessed data as input; constructing a NetTDF model based on a U-shaped network, fusing the features extracted by the MTH3 model, and modeling a global relationship; and dividing a training set and a test set, and inputting the training set and the test set into the overall model to obtain a prediction result of the lithium ion battery health state. According to the method provided by the invention, the advantages of each model are fully exerted, an efficient prediction model for the capacity of the lithium ion battery, which can accurately capture local features and effectively model a global relationship, is provided, and reliable support is provided for performance evaluation of the lithium ion battery.

Description

Technical Field

[0001] The present invention relates to the technical field of lithium ion battery performance evaluation, and in particular to a time series modeling method for lithium ion battery performance evaluation. Background Art

[0002] With the rapid advancement of new energy technologies, lithium-ion batteries have been widely used in electric vehicles, portable electronic devices, energy storage systems and other fields. The performance status of the battery directly affects the operating efficiency and safety of the equipment. Therefore, it is crucial to accurately assess the health of the battery. Battery health status is the core indicator for evaluating battery performance. Traditional health status prediction methods usually rely on physical models or empirical formulas, but these methods often require a lot of experimental data and professional technical support, and show great limitations when facing the behavior of batteries in complex and changing working environments.

[0003] In recent years, data-driven battery health status prediction methods have gradually become a research hotspot. These methods automatically extract key features from a large amount of battery operation data by using machine learning and deep learning techniques, and build prediction models based on these features. Among the many data-driven methods, the traditional Transformer model and temporal convolutional network (TCN) have been widely used due to their outstanding performance in processing time series data. Transformer effectively captures the global dependencies in the sequence through the multi-head self-attention mechanism, but is slightly insufficient in processing local time series features. TCN is particularly outstanding in capturing short-term changes and local time series features, and is particularly good at processing these local information. However, due to the locality of the convolution operation, TCN has certain limitations in capturing global features and it is difficult to fully model long-term dependencies. In recent years, some new neural network models such as the Mamba model, with its hierarchical structure and forgetting mechanism, can automatically extract key time series features from complex battery data, showing significant advantages. It may be possible to consider introducing the Mamba model to make up for the shortcomings of the Transformer model and the temporal convolutional TCN model. Summary of the invention

[0004] In view of the shortcomings of the prior art, the present invention provides a time series modeling method for lithium-ion battery performance evaluation, solves the technical problems of traditional methods in multi-scale feature extraction and fusion, and can improve the accuracy and robustness of battery health status prediction; the present invention combines the hierarchical structure and forgetting mechanism of the Mamba model, innovatively integrates the temporal convolutional network (TCN) and the multi-head self-attention mechanism (MHSA), and proposes a multi-module architecture-MTH (MambaTCN Heads). This architecture can more effectively process and model time series data, and improve the performance and adaptability of the model on time series data. In order to more effectively fuse features of different time scales, the present invention introduces the time series fusion model NetTDF (Unet-based Temporal dimension fusion) based on the U-net network U-Net. The U-net structure is used to more effectively capture global features and realize the fusion of multi-scale features, thereby enhancing the model's comprehensive understanding of information at different time scales. Through this innovative solution, the present invention gives full play to the advantages of each model, aiming to provide an efficient prediction model that can accurately capture local features and effectively model global relationships.

[0005] In order to achieve the above technical objectives, the present invention provides the following technical solutions:

[0006] The time series modeling method for lithium-ion battery performance evaluation specifically includes the following steps:

[0007] S1. Obtain the original capacity data of lithium-ion batteries from public data sets and preprocess the original capacity data;

[0008] S2, based on the structural design and forgetting mechanism of the Mamba model, the MTH module is constructed by combining the temporal convolutional network TCN and the multi-head attention mechanism MHSA; then the MTH3 model is obtained by stacking three MTH modules, and the preprocessed battery capacity data is used as the input of the MTH3 model to extract data features at different time scales;

[0009] S3, introduce the time series fusion model NetTDF based on U-type network, take the data features of different time scales extracted by MTH3 model as input, and perform multi-scale feature fusion;

[0010] S4. Divide the preprocessed lithium-ion battery capacity data into a training set and a test set, train and test the MTH3 model and the NetTDF model, and obtain the prediction results of the health status of the lithium-ion battery.

[0011] Furthermore, step S1 specifically includes:

[0012] S11. Obtaining original capacity data X and rated capacity data C0 of a lithium-ion battery from a public data set; the public data set is a lithium-ion battery data set provided by NASA;

[0013] S12, divide the original capacity data by the rated capacity data to normalize, and obtain the normalized capacity data X normalized ;

[0014] S13, regularize the normalized capacity data again, reshape it into the shape required by the MTH3 model and the NetTDF model, and complete the preprocessing; the preprocessed battery capacity data X re As input to the MTH3 model.

[0015] Furthermore, the MTH3 model in step S2 is specifically:

[0016] Each MTH module specifically includes a linear transformation module, a TCN temporal convolution module, an activation function ReLU, a multi-head self-attention module MHSA, and an element-by-element multiplication module connected in sequence; at the same time, an activation function Sigmoid is added after the linear transformation module as a forgetting mechanism, and the output obtained by the forgetting mechanism is multiplied element-by-element with the output of the MHSA module to obtain the final output of each MTH module;

[0017] The MTH3 model is obtained by stacking three MTH modules with the same structure.

[0018] Furthermore, the process of extracting data features at different time scales in step S2 specifically includes:

[0019] S21, the pre-processed battery capacity data X re They are input to each MTH module respectively. Each MTH module first projects the time dimension to a low dimension with different time steps through a linear transformation module to obtain low-dimensional time features at different time scales, which are recorded as X 1-1 , X 2-1 , X 3-1 ;

[0020] S22. In each MTH module, the low-dimensional time feature X 1-1 , X 2-1 , X 3-1 Then the output X is obtained through the TCN time convolution module respectively. 1-2 , X 2-2 , X 3-2 ;

[0021] S23, the output X of the temporal convolution module 1-2 , X 2-2 , X 3-2First, it is transformed into X through an activation function ReLU 1-2 , X 2-2 , X 3-2 , and then reshape their shapes into the shapes required by the multi-head self-attention module MHSA through the reshaping operation, that is, the initial input X1′ of MHSA is obtained -3 , X2′ -3 , X3′ -3 ;

[0022] S24. In the multi-head self-attention module MHSA, set 4 attention heads i=0, 1, 2, 3, and use different weight matrices W Q (i) , W K (i) and W V (i) For X1′ -3 , X2′ -3 , X3′ -3 Generate query matrix Q i , key matrix K i Sum value matrix V i ; Then use the generated query matrix, key matrix and value matrix to calculate the scaled dot product attention head of different attention heads i , then concatenate the outputs of all heads and pass them through a linear transformation matrix W o The concatenated result is mapped to the final output space, and then the output X1″ of the MHSA module is obtained through residual connection and layer normalization. -3 , X2″ -3 , X3″ -3 ;

[0023] S25, the output of the MHSA module is X1″ -3 , X2″ -3 , X3″ -3 Reshape it back to the shape before inputting into MHSA, and record the output as X after reshaping again 1-4 ,X 2-4 ,X 3-4 ;

[0024] S26, add a forgetting mechanism; obtain the low-dimensional time features X at different time scales in step S21 1-1 , X 2-1 , X 3-1 Through an activation function Sigmoid, we get the output X after the forgetting mechanism is processed. 1-5 ,X 2-5 ,X 3-5 ;

[0025] S27, X 1-4 ,X 2-4 ,X3-4 and X 1-5 ,X 2-5 ,X 3-5 Perform element-by-element multiplication to obtain the final output X1, X2, and X3 of each MTH module, that is, the MTH3 model.

[0026] Furthermore, step S3 specifically includes the following steps:

[0027] S31, taking the final output of the MTH3 module as input, perform feature fusion based on the NetTDF model to obtain the final output fusion feature, recorded as Y';

[0028] S32, Y' and the normalized capacity data X normalized Through the residual connection operation, the obtained feature is recorded as Y″;

[0029] S33. Perform a linear transformation on Y″ to obtain the final output Y of the time series model MTH3-NetTDF.

[0030] More specifically, step S31 is as follows:

[0031] Taking the final output of the MTH3 module as input, the final outputs X1, X2, and X3 of the MTH3 module are concatenated in the time dimension from bottom to top; that is, the output X3 of the bottom layer is concatenated with the output X2 of the middle layer first, and the concatenated features are then passed through a linear transformation module to obtain the concatenated output X2′ of the middle layer, and then the concatenated output of the middle layer is concatenated with the output X2′ and X1 of the top layer, and the concatenated features are then passed through a linear layer to obtain the final output fusion feature Y' of the NetTDF model.

[0032] Furthermore, step S4 specifically includes:

[0033] S41, the pre-processed lithium-ion battery data is divided into a training set and a test set; the training set and the test set are further divided according to the batch size B; the divided training set is first input into the MTH3 model and the NetTDF model in batches for training; after multiple trainings, the optimal MTH3-NetTDF model is obtained;

[0034] S42. Then, the test set is input into the trained optimal model to obtain the evaluation result of the lithium-ion battery performance.

[0035] By means of the above technical solution, the present invention proposes a time series modeling method for lithium-ion battery performance evaluation, which has at least the following beneficial effects:

[0036] Compared with traditional methods, the present invention proposes an innovative architecture of multi-module combination to improve the processing capability and prediction accuracy of time series data. Specifically, the present invention combines the Mamba model, TCN model and multi-head self-attention mechanism to extract features for data of different time scales, significantly improving the performance of the overall model in processing local time series features; at the same time, the U-type network-based time series fusion model NetTDF is introduced, which captures global features more effectively through the U-type network and realizes the fusion of multi-scale features, further enhancing the model's comprehensive understanding of information at different time scales; through the multi-model combination scheme designed by the present invention, the advantages of each model are fully utilized, providing an efficient prediction model for lithium-ion battery capacity that can accurately capture local features and effectively model global relationships, providing reliable support for lithium-ion battery performance evaluation. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0038] Figure 1 It is the overall flow chart of the method proposed by the present invention;

[0039] Figure 2 A schematic diagram of the structure of the MTH3-NetTDF model designed for the present invention;

[0040] Figure 3 The single-step and multi-step prediction diagrams of the health status of battery B0005 in the public dataset;

[0041] Figure 4 The single-step and multi-step prediction diagrams of the health status of battery B0006 in the public dataset;

[0042] Figure 5 This is the single-step and multi-step prediction diagram of the health status of battery B0007 in the public dataset;

[0043] Figure 6 Single-step and multi-step prediction diagrams of the health status of battery B0018 in the public dataset;

[0044] Figure 7 The following is a comparison chart of the prediction effects of different models based on three evaluation indicators. DETAILED DESCRIPTION

[0045] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0046] Although the steps in the present invention are arranged with numbers, they are not used to limit the order of the steps. Unless the order of the steps is clearly stated or the execution of a certain step requires other steps as a basis, the relative order of the steps can be adjusted. It is understood that the term "and / or" used in this article involves and covers any and all possible combinations of one or more of the associated listed items.

[0047] Please refer to Figure 1 , showing the time series modeling method for lithium-ion battery performance evaluation proposed in the present invention, which specifically includes the following steps:

[0048] S1. Obtain the original capacity data of lithium-ion batteries from public data sets and preprocess the original capacity data;

[0049] As a preferred implementation, step S1 specifically includes:

[0050] S11. Obtaining original capacity data X and rated capacity data C0 of a lithium-ion battery from a public data set; the public data set is a lithium-ion battery data set provided by NASA;

[0051] S12, divide the original capacity data by the rated capacity data, and normalize the data to be within the range of 0-1 to obtain the normalized capacity data X normalized ; The formula is:

[0052]

[0053] S13, regularize the normalized capacity data again, reshape it into the shape required by the MTH3 model and the NetTDF model, and complete the preprocessing; the preprocessed battery capacity data X re As the input of the MTH3 model; the formula of this process is expressed as:

[0054]

[0055] in, Represents X re is a real number tensor of shape C×N×T; (Similar expressions elsewhere in the text are also expressed as real number tensors of corresponding shapes, which will not be repeated in this embodiment)

[0056] C represents the number of features, N represents the number of sequences, T represents the time step, and reshape(·) represents the reshaping operation.

[0057] S2, based on the structural design and forgetting mechanism of the Mamba model, the MTH module is constructed by combining the temporal convolutional network TCN and the multi-head attention mechanism MHSA; then the MTH3 model is obtained by stacking three MTH modules, and the preprocessed battery capacity data is used as the input of the MTH3 model to extract data features at different time scales;

[0058] Each MTH module designed by the present invention specifically includes a linear transformation module, a TCN time convolution module, an activation function ReLU, a multi-head self-attention module MHSA and an element-by-element multiplication module connected in sequence; at the same time, an activation function Sigmoid is added after the linear transformation module as a forgetting mechanism, and the output obtained by the forgetting mechanism is multiplied element by element with the output of the MHSA module to obtain the final output of each MTH module;

[0059] The MTH3 model was obtained by stacking three MTH modules with the same structure;

[0060] Based on the above MTH3 model, as a preferred implementation, the process of extracting data features at different time scales in step S2 specifically includes:

[0061] S21, the pre-processed battery capacity data X re They are input to each MTH module respectively. Each MTH module first projects the time dimension to a low dimension with different time steps through a linear transformation module to obtain low-dimensional time features at different time scales, which are recorded as X 1-1 , X 2-1 , X 3-1 ; The formula is:

[0062]

[0063] Wherein, Linear(·) represents a linear transformation; in this embodiment, the time step of the first MTH module is consistent with the input preprocessed battery capacity data, and the time step of the second and third MTH modules is reduced by 4 based on the previous one, so that low-dimensional time features at different time scales are obtained;

[0064] S22. In each MTH module, the low-dimensional time feature X 1-1 , X 2-1 , X 3-1 Then the output X is obtained through the TCN time convolution module respectively. 1-2 , X 2-2 , X 3-2 ;TCN stacks multiple convolutional layers to further capture temporal dependencies at different levels;

[0065] S23, the output X of the temporal convolution module1-2 , X 2-2 , X 3-2 First, it is transformed into X through an activation function ReUL 1-2 , X 2-2 , X 3-2 , the formula is:

[0066]

[0067] In this embodiment, the activation function ReLU introduces nonlinearity, allowing the model to learn more complex, nonlinear features; helping the model to better fit the training data and improve the generalization ability of the network

[0068] Then, through the reshaping operation, each shape is reshaped into the shape required by the multi-head self-attention module MHSA, that is, the initial input X1′ of MHSA is obtained -3 , X2′ -3 , X3′ -3 ; The formula is:

[0069]

[0070] S24. In the multi-head self-attention module MHSA, set 4 attention heads i=0, 1, 2, 3, and use different weight matrices W Q (i) , W K (i) and W V (i) For X1′ -3 , X2′ -3 , X3′ -3 Generate query matrix Q i , key matrix K i Sum value matrix V i ; Then use the generated query matrix, key matrix and value matrix to calculate the scaled dot product attention head of different attention heads i , then concatenate the outputs of all heads and pass them through a linear transformation matrix W o The concatenated result is mapped to the final output space, and then the output X1″ of the MHSA module is obtained through residual connection and layer normalization. -3 , X2″ -3 , X3″ -3 ;

[0071] Here X1′ -3 For input X1″ -3 Taking the output as an example, the formula of the processing process of the MHSA module is expressed as:

[0072]

[0073] Among them, d k represents the dimension of the key, represents the scaling factor, softmax(·) represents the conversion of similarity into attention weight, Attention(·) represents the attention calculation; Concat(·) represents the concatenation operation;

[0074] In this application, MHSA captures global information in parallel through multiple attention heads, and can more efficiently model long-term dependencies in long time series; compared with the SSM module in the original mamba model, MHSA can be calculated in parallel through the attention mechanism, and does not need to recursively update the state step by step, which makes it more computationally efficient when processing long time series; in addition, by adjusting the number and number of attention heads, it can flexibly capture the relationship between different time periods and adapt to more complex tasks;

[0075] S25, the output of the MHSA module is X1″ -3 , X2″ -3 , X3″ -3 Reshape it back to the shape before inputting into MHSA, and record the output as X after reshaping again 1-4 ,X 2-4 ,X 3-4 ; The formula is:

[0076]

[0077] S26, add a forgetting mechanism; obtain the low-dimensional time features X at different time scales in step S21 1-1 , X 2-1 , X 3-1 Through an activation function Sigmoid, we get the output X after the forgetting mechanism is processed. 1-5 ,X 2-5 ,X 3-5 ; The formula is:

[0078]

[0079] Among them, σ(.) represents the activation function Sigmoid;

[0080] S27, X 1-4 ,X 2-4 ,X 3-4 and X 1-5 ,X 2-5 ,X 3-5 Perform element-by-element multiplication to obtain the final output X1, X2, X3 of each MTH module, that is, the MTH3 model; the formula is expressed as:

[0081]

[0082] in, Represents element-wise multiplication;

[0083] In this embodiment, the activation function Sigmoid generates a weight between 0 and 1 according to the input features and the current time step. The weight is multiplied element by element with the output of MHSA to control the selective memory and discarding of the output information; that is, in the training process, the forgetting mechanism is used to dynamically determine which information should be "forgotten" and which information should be retained, so as to better adapt to the needs of the current task.

[0084] S3, introduce the time series fusion model NetTDF based on U-type network, take the data features of different time scales extracted by MTH3 model as input, and perform multi-scale feature fusion;

[0085] As a preferred implementation, step S3 specifically includes:

[0086] S31, taking the final output of the MTH3 module as input, perform feature fusion based on the NetTDF model to obtain the final output fusion feature, recorded as Y';

[0087] More specifically, Figure 2 As shown in , step S31 is specifically as follows:

[0088] Taking the final output of the MTH3 module as input, the final outputs X1, X2, and X3 of the MTH3 module are concatenated in the time dimension from bottom to top; that is, the output X3 of the bottom layer is concatenated with the output X2 of the middle layer, and the concatenated features are passed through a linear transformation module to obtain the concatenated output X2′ of the middle layer, and then the concatenated output of the middle layer is concatenated with the output X2′ and X1 of the top layer, and the concatenated features are passed through a linear layer to obtain the final output fusion feature Y' of the NetTDF model;

[0089] The formula is:

[0090]

[0091] Among them, Concat(·) represents the time dimension concatenation operation;

[0092] S32, Y' and the normalized capacity data X normalized Through the residual connection operation, the feature obtained is recorded as Y″; the formula is expressed as:

[0093]

[0094] S33, perform a linear transformation on Y″ to obtain the final output Y of the time series model MTH3-NetTDF; the formula is expressed as:

[0095]

[0096] The residual connection in step S32 above directly connects the input X normalized Adding it to the output Y' ensures that some original information in the network will not be completely discarded, reducing information loss; and the linear transformation in step S33 maps Y" to the final output space, so that actual prediction can be performed.

[0097] It should also be noted that the time series model MTH3-NetTDF that appears in this embodiment refers to the lithium-ion battery capacity prediction model designed by the present invention, which processes the time series data of battery capacity data. MTH3-NetTDF represents two core models.

[0098] S4, dividing the preprocessed lithium-ion battery capacity data into a training set and a test set, training and testing the MTH3 model and the NetTDF model, and obtaining the prediction results of the health status of the lithium-ion battery;

[0099] As a preferred implementation, step S4 specifically includes:

[0100] S41, the pre-processed lithium-ion battery data is divided into a training set and a test set; the training set and the test set are further divided according to the batch size B; the divided training set is first input into the MTH3 model and the NetTDF model in batches for training; after multiple trainings, the optimal MTH3-NetTDF model is obtained;

[0101] S42, then input the test set into the trained optimal model. In this embodiment, the battery capacity data in the test set is recorded as The optimal model obtained is expressed as MN(·), and the test result is That is, the evaluation result of the lithium-ion battery performance is obtained.

[0102] This embodiment also provides the following experimental examples to verify the superiority of the method proposed in the present invention:

[0103] In the experimental example, the "leave one out" method was used for training and testing. Specifically, through four rounds of iterations, the data of one battery was retained for testing each time, and finally all four batteries were tested. All data were normalized and mapped to the range of [0,1] by dividing by the rated capacity of the battery (Rated_Capacity=2.0). Subsequently, the normalized data was split according to the batch size (batch_size=16) and reshaped into a tensor of shape B×C×N×T, which was input into the MTH3-NetTDF model. The model generates more accurate health status prediction results by extracting multi-scale features in the time series.

[0104] In the specific experiment, the lithium battery capacity data provided by NASA was used, and four batteries (No. 5, 6, 7 and 18) were selected from the 18650 model. The experiment was carried out at room temperature (24°C), and three modes of charging, discharging and electrochemical impedance spectroscopy (EIS) were operated respectively. When charging, a constant current (CC) mode of 1.5A was used until the battery voltage reached 4.2V, and then switched to a constant voltage (CV) mode and continued charging until the current dropped to 20mA. During the discharge process, a constant current (CC) mode of 2A was used until the battery voltage dropped to a predetermined different value: No. 5, No. 6, and No. 7 batteries dropped to 2.7V, 2.5V, and 2.2V, respectively, and No. 18 battery dropped to 2.5V; the EIS measurement frequency range was 0.1Hz to 5kHz. In order to evaluate the effect of EIS on capacity, 278 EIS measurements were performed on No. 5, No. 6 and No. 7 batteries, and 53 on No. 18 battery. The experiment was stopped when the battery reached the end of life (EOL), specifically when the rated capacity of the battery dropped by 30% (from 2Ah to 1.4Ah). The entire experiment consisted of 616 small processes and 168 or 132 complete cycles.

[0105] The experimental results are as follows Figures 3 to 6 As shown, the performance of the method of the present invention under single-step prediction and multi-step prediction is demonstrated. Although the effects of the two prediction strategies on different batteries are different, they both reflect the superiority of the present invention in estimating the health status of lithium-ion batteries. The single-step prediction results show that the method is highly consistent with the actual situation in short-term prediction, while the multi-step prediction results show good long-term prediction accuracy and reliability.

[0106] Furthermore, the present invention also compares the proposed MTH3-NetTDF model with five baseline methods in the prior art (MLP, RNN, LSTM, GRU, Dual-LSTM and DeTransformer) in terms of lithium-ion battery capacity prediction. The comparison results are shown in Tables 1 and Figure 7 As shown;

[0107] Table 1 Comparison results between this model and existing models

[0108]

[0109]

[0110] By evaluating indicators such as relative error (RE), mean absolute error (MAE) and root mean square error (RMSE), the results show that the proposed model outperforms other baseline methods in all evaluation indicators, indicating that this method can more accurately capture the changing trend of the battery health status and thus provide more accurate capacity prediction.

[0111] In summary, the present invention innovatively introduces TCN and MHSA to replace the traditional convolutional module and SSM module respectively based on the hierarchical structure and forgetting mechanism of the Mamba model, and proposes an MTH module architecture. This architecture significantly improves the performance of the model in capturing local and global time series features through multi-scale feature extraction and time series fusion. At the same time, the introduction of the U-network-based time series fusion model NetTDF further optimizes the fusion process of multi-scale features, allowing the model to better understand complex time series data; compared with the prior art, the method proposed by the present invention is more accurate in battery capacity prediction, and can better capture the time series change characteristics of the battery health status, providing a more reliable technical option for lithium-ion battery performance evaluation.

[0112] The terms used in this specification, such as "one embodiment", "some embodiments", "example", "specific example" or "some examples", etc., all refer to the features, structures, materials or characteristics involved in a specific implementation or example, which may appear in multiple embodiments or examples of the present invention and can be combined as needed. In addition, those skilled in the art can reasonably combine the different embodiments or examples and their features described in this specification without causing conflicts.

[0113] The logic and steps shown in the flowchart, or the implementation described in other ways, can be regarded as a sequence of executable instructions for implementing specific functions. These instructions can be implemented in any computer-readable medium and used by an instruction execution system, device or apparatus (for example, a computer system, a system including a processor, or any system that can obtain and execute instructions from an instruction execution system, device or apparatus), or used in combination with these systems, devices or apparatuses.

[0114] The embodiments of the present invention have been described in detail in the above content, and the principles and embodiments of the present invention have been described by specific examples. The above embodiments are only for understanding the method of the present invention and its core concept. For those skilled in the art, according to the concept of the present invention, the specific embodiments and the scope of application may be different. Therefore, the description herein should not be regarded as limiting the present invention.

Claims

1. A time series modeling method for lithium-ion battery performance evaluation, characterized in that: The specific steps include: S1. Obtain the original capacity data of lithium-ion batteries from public data sets and preprocess the original capacity data; S2, based on the structural design and forgetting mechanism of the Mamba model, the MTH module is constructed by combining the temporal convolutional network TCN and the multi-head attention mechanism MHSA; Then, the MTH3 model is obtained by stacking three MTH modules. The preprocessed battery capacity data is used as the input of the MTH3 model to extract data features at different time scales. S3, introduce the time series fusion model NetTDF based on U-type network, take the data features of different time scales extracted by MTH3 model as input, and perform multi-scale feature fusion; S4. Divide the preprocessed lithium-ion battery capacity data into a training set and a test set, train and test the MTH3 model and the NetTDF model, and obtain the prediction results of the health status of the lithium-ion battery.

2. The time series modeling method for lithium-ion battery performance evaluation according to claim 1, characterized in that: Step S1 specifically includes: S11. Obtaining original capacity data X and rated capacity data C0 of a lithium-ion battery from a public data set; the public data set is a lithium-ion battery data set provided by NASA; S12, divide the original capacity data by the rated capacity data to normalize, and obtain the normalized capacity data X normalized ; S13, regularize the normalized capacity data again, reshape it into the shape required by the MTH3 model and the NetTDF model, and complete the preprocessing; the preprocessed battery capacity data X re As input to the MTH3 model.

3. The time series modeling method for lithium-ion battery performance evaluation according to claim 1, characterized in that: The MTH3 model in step S2 is specifically: Each MTH module specifically includes a linear transformation module, a TCN temporal convolution module, an activation function ReLU, a multi-head self-attention module MHSA, and an element-by-element multiplication module connected in sequence; at the same time, an activation function Sigmoid is added after the linear transformation module as a forgetting mechanism, and the output obtained by the forgetting mechanism is multiplied element-by-element with the output of the MHSA module to obtain the final output of each MTH module; The MTH3 model is obtained by stacking three MTH modules with the same structure.

4. The time series modeling method for lithium-ion battery performance evaluation according to claim 3, characterized in that: The process of extracting data features at different time scales in step S2 specifically includes: S21, the pre-processed battery capacity data X re They are input to each MTH module respectively. Each MTH module first projects the time dimension to a low dimension with different time steps through a linear transformation module to obtain low-dimensional time features at different time scales, which are recorded as X 1-1 , X 2-1 , X 3-1 ; S22. In each MTH module, the low-dimensional time feature X 1-1 , X 2-1 , X 3-1 Then the output X is obtained through the TCN time convolution module respectively. 1-2 , X 2-2 , X 3-2 ; S23, the output X of the temporal convolution module 1-2 , X 2-2 , X 3-2 First, it is transformed into X through an activation function ReLU 1-2 , X 2-2 , X 3-2 , and then reshape their shapes into the shapes required by the multi-head self-attention module MHSA through the reshaping operation, that is, the initial input X1′ of MHSA is obtained -3 , X2′ -3 , X3′ -3 ; S24. In the multi-head self-attention module MHSA, set 4 attention heads i=0, 1, 2, 3, and use different weight matrices and For X1′ -3 , X2′ -3 , X3′ -3 Generate query matrix Q i , key matrix K i Sum value matrix V i ; Then use the generated query matrix, key matrix and value matrix to calculate the scaled dot product attention head of different attention heads i , then concatenate the outputs of all heads and pass them through a linear transformation matrix W o The concatenated result is mapped to the final output space, and then the output X1″ of the MHSA module is obtained through residual connection and layer normalization. -3 , X2″ -3 , X3″ -3 ; S25, the output of the MHSA module is X1″ -3 , X2″ -3 , X3″ -3 Reshape it back to the shape before inputting into MHSA, and record the output as X after reshaping again 1-4 ,X 2-4 ,X 3-4 ; S26, add a forgetting mechanism; obtain the low-dimensional time features X at different time scales in step S21 1-1 , X 2-1 , X 3-1 Through an activation function Sigmoid, we get the output X after the forgetting mechanism is processed. 1-5 ,X 2-5 ,X 3-5 ; S27, X 1-4 ,X 2-4 ,X 3-4 and X 1-5 ,X 2-5 ,X 3-5 Perform element-by-element multiplication to obtain the final output X1, X2, and X3 of each MTH module, that is, the MTH3 model.

5. The time series modeling method for lithium-ion battery performance evaluation according to claim 2, characterized in that: Step S3 specifically includes the following steps: S31, taking the final output of the MTH3 module as input, perform feature fusion based on the NetTDF model to obtain the final output fusion feature, recorded as Y'; S32, Y' and the normalized capacity data X normalized Through the residual connection operation, the obtained feature is recorded as Y″; S33. Perform a linear transformation on Y″ to obtain the final output Y of the time series model MTH3-NetTDF.

6. The time series modeling method for lithium-ion battery performance evaluation according to claim 5, characterized in that: Step S31 is specifically as follows: Taking the final output of the MTH3 module as input, the final outputs X1, X2, and X3 of the MTH3 module are concatenated in the time dimension from bottom to top; that is, the output X3 of the bottom layer is concatenated with the output X2 of the middle layer first, and the concatenated features are then passed through a linear transformation module to obtain the concatenated output X2′ of the middle layer, and then the concatenated output of the middle layer is concatenated with the output X2′ and X1 of the top layer, and the concatenated features are then passed through a linear layer to obtain the final output fusion feature Y' of the NetTDF model.

7. The time series modeling method for lithium-ion battery performance evaluation according to claim 1, characterized in that: Step S4 specifically includes: S41, the pre-processed lithium-ion battery data is divided into a training set and a test set; the training set and the test set are further divided according to the batch size B; the divided training set is first input into the MTH3 model and the NetTDF model in batches for training; after multiple trainings, the optimal MTH3-NetTDF model is obtained; S42. Then, the test set is input into the trained optimal model to obtain the evaluation result of the lithium-ion battery performance.

Images