Battery Pack State of Health Prediction Method and Device

Abstract

The present disclosure relates to a method and device for predicting the state of health of a battery pack, including: obtaining historical time-series operation data of each single battery in the battery pack to be predicted, extracting features from the historical time-series operation data to obtain a feature matrix of the battery pack; constructing a static adjacency graph, a health characteristic graph, and a historical dependence graph of the battery pack based on the historical time-series operation data; inputting the feature matrix and the static adjacency graph into a dynamic graph generation network of a state-of-health prediction model to obtain a dynamic adjacency graph; inputting the multi-feature graph into a graph convolutional network of the state-of-health prediction model to obtain spatial features; and inputting the spatial features into a KAN network of the state-of-health prediction model to obtain state-of-health prediction values of each single battery in the battery pack. Based on the trained state-of-health prediction model, it can effectively cope with the high-dimensional non-linear complexity of the operation characteristics of the battery pack and can more accurately and efficiently predict the state-of-health values of each single battery in the battery pack.

Description

Technical Field

[0001] The present disclosure relates to the field of new energy technologies, and particularly to a method and device for predicting the health state of a battery pack. Background Art

[0002] Lithium-ion batteries have become an important part of modern energy systems due to their high energy density, long lifespan, good cycling performance, and environmental friendliness, and are widely used in energy storage power stations, electric vehicles, smart grids, and other fields. However, with the increase in usage time and the number of cycles, the performance of lithium-ion batteries inevitably degrades gradually. Capacity attenuation and lifespan reduction will directly affect the operating efficiency and system safety of the battery pack. Especially in multi-battery systems such as energy storage power stations, the performance degradation of individual batteries may even trigger a chain reaction, thus threatening the stability and safety of the entire battery pack. Therefore, how to accurately predict the health state of individual batteries in a battery pack has become an urgent problem to be solved. Summary of the Invention

[0003] The present disclosure provides a method, device, electronic device, and computer-readable storage medium for predicting the health state of a battery pack.

[0004] The technical solution of the present disclosure is as follows:

[0005] According to the first aspect of the embodiments of the present disclosure, a method for predicting the health state of a battery pack is provided. The method includes: obtaining historical sequential operation data of each individual battery in the battery pack to be predicted, performing feature extraction on the historical sequential operation data to obtain a feature matrix of the battery pack; constructing a static adjacency graph, a health characteristic graph, and a historical dependence graph of the battery pack based on the historical sequential operation data; inputting the feature matrix and the static adjacency graph into a dynamic graph generation network of a health state prediction model to obtain a dynamic adjacency graph output by the dynamic graph generation network; inputting a multi-feature graph into a graph convolutional network of the health state prediction model to obtain spatial features output by the graph convolutional network, where the multi-feature graph includes the static adjacency graph, the health characteristic graph, the historical dependence graph, and the dynamic adjacency graph; and inputting the spatial features into a KAN network of the health state prediction model to obtain health state prediction values of each individual battery in the battery pack output by the KAN network.

[0006] According to a second aspect of the embodiments of the present disclosure, a battery pack state of health prediction device is provided, including: a first acquisition module, configured to acquire historical sequential operation data of each single battery in the battery pack to be predicted, perform feature extraction on the historical sequential operation data, and acquire a feature matrix of the battery pack; a second acquisition module, configured to construct a static adjacency graph, a health characteristic graph, and a historical dependence graph of the battery pack based on the historical sequential operation data; a first processing module, configured to input the feature matrix and the static adjacency graph into a dynamic graph generation network of a state of health prediction model to acquire a dynamic adjacency graph output by the dynamic graph generation network; a second processing module, configured to input a multi-feature graph into a graph convolutional network of the state of health prediction model to acquire spatial features output by the graph convolutional network, where the multi-feature graph includes the static adjacency graph, the health characteristic graph, the historical dependence graph, and the dynamic adjacency graph; a third processing module, configured to input the spatial features into a KAN network of the state of health prediction model to acquire state of health prediction values of each single battery in the battery pack output by the KAN network.

[0007] According to a third aspect of the embodiments of the present disclosure, an electronic device is provided, including: a processor; a memory for storing executable instructions of the processor; wherein, the processor is configured to execute the instructions to implement the battery pack state of health prediction method as described in the first aspect of the embodiments of the present disclosure.

[0008] According to a fourth aspect of the embodiments of the present disclosure, a computer-readable storage medium is provided, which, when instructions in the computer-readable storage medium are executed by a processor of an electronic device, enables the electronic device to execute the battery pack state of health prediction method as described in the first aspect of the embodiments of the present disclosure.

[0009] The technical solutions provided by the embodiments of the present disclosure at least bring the following beneficial effects:

[0010] In an embodiment of the present disclosure, by obtaining the historical time-series operation data of each single battery in the battery pack to be predicted, extracting features from the historical time-series operation data to obtain the feature matrix of the battery pack, constructing a static adjacency graph, a health characteristic graph, and a historical dependence graph of the battery pack based on the historical time-series operation data, inputting the feature matrix and the static adjacency graph into the dynamic graph generation network of the health state prediction model to obtain the dynamic adjacency graph output by the dynamic graph generation network, inputting the multi-feature graph into the graph convolutional network of the health state prediction model to obtain the spatial features output by the graph convolutional network, where the multi-feature graph includes the static adjacency graph, the health characteristic graph, the historical dependence graph, and the dynamic adjacency graph, inputting the spatial features into the KAN network of the health state prediction model to obtain the health state prediction values of each single battery in the battery pack output by the KAN network. Thus, based on the trained health state prediction model, the present disclosure can effectively cope with the high-dimensional non-linear complexity of the operation characteristics of the battery pack, and can accurately and efficiently predict the health state values of each single battery in the battery pack, which is beneficial to improving the operation safety and stability of the battery system.

[0011] It should be understood that the above general description and the following detailed description are only exemplary and explanatory, and do not limit the present disclosure. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] The accompanying drawings herein are incorporated into the specification and constitute a part of the specification, showing embodiments consistent with the present disclosure, and are used together with the specification to explain the principles of the present disclosure, and do not constitute an improper limitation to the present disclosure.

[0013] Figure 1 is a flowchart of a method for predicting the health state of a battery pack shown according to an exemplary embodiment.

[0014] Figure 2 is a flowchart of a method for predicting the health state of a battery pack shown according to another exemplary embodiment.

[0015] Figure 3 is a flowchart of a method for predicting the health state of a battery pack shown according to another exemplary embodiment.

[0016] Figure 4 is a flowchart of a method for predicting the health state of a battery pack shown according to another exemplary embodiment.

[0017] Figure 5 is a flowchart of a method for predicting the health state of a battery pack shown according to another exemplary embodiment.

[0018] Figure 6 is a flowchart of a method for predicting the health state of a battery pack shown according to another exemplary embodiment.

[0019] Figure 7 It is a block diagram of a state of health prediction device for a battery pack shown according to another exemplary embodiment.

[0020] Figure 8 It is a block diagram of an electronic device shown according to another exemplary embodiment. Detailed implementation manners

[0021] In order to enable those of ordinary skill in the art to better understand the technical solutions of the present disclosure, the technical solutions in the embodiments of the present disclosure will be clearly and completely described below with reference to the accompanying drawings.

[0022] It should be noted that the terms "first", "second", etc. in the specification of the present disclosure and the above-mentioned drawings are used to distinguish similar objects, and do not necessarily need to describe a specific order or sequence. It should be understood that such used data can be interchanged under appropriate circumstances so that the embodiments of the present disclosure described herein can be implemented in an order other than those illustrated or described herein. The implementation manners described in the following exemplary embodiments do not represent all implementation manners consistent with the present disclosure.

[0023] Hereinafter, embodiments are used to elaborate in detail on the battery pack state of health prediction method, device, and electronic device proposed by the present disclosure.

[0024] Figure 1 It is a schematic flowchart of a battery pack state of health prediction method provided for an embodiment of the present disclosure.

[0025] As Figure 1 shown, the battery pack state of health prediction method proposed in this embodiment includes the following steps:

[0026] S101. Obtain the historical time-series operation data of each single battery in the battery pack to be predicted, perform feature extraction on the historical time-series operation data, and obtain the feature matrix of the battery pack.

[0027] Optionally, the historical time-series operation data of each single battery in the battery pack to be predicted can be obtained from the historical time-series operation database of the battery pack.

[0028] In the embodiment of the present disclosure, after obtaining the historical time-series operation data, feature extraction can be performed on the historical time-series operation data to obtain the feature matrix of the battery pack, such as: various feature matrices such as the state of charge (SOC) range, charging capacity, voltage range, etc.

[0029] S102. Based on the historical time-series operation data, construct a static adjacency graph, a health characteristic graph, and a historical dependency graph of the battery pack.

[0030] In the embodiments of the present disclosure, based on historical sequential operation data, the temperature, state of charge, and initial capacity of every two single cells in the battery pack can be obtained, and a static adjacency graph of the battery pack can be constructed based on the temperature, state of charge, and initial capacity of every two single cells.

[0031] Optionally, each element in the static adjacency matrix of the battery pack can be determined based on the temperature, state of charge, and initial capacity of every two single cells, and a static adjacency matrix can be constructed based on each element in the static adjacency matrix of the battery pack, where each element is used to describe the static interaction relationship between every two single cells, and the static adjacency matrix is converted into a static adjacency graph. 。

[0032] It should be noted that traditional static adjacency graphs are usually constructed based on fixed physical connections or simple temperature and current relationships. To improve the expression ability of the static adjacency graph for static interaction relationships, the present disclosure introduces a hybrid static mapping.

[0033] For example, for single cells in the battery pack i and single cell j , the element in the static adjacency matrix of the battery pack can be determined based on the temperature, state of charge, and initial capacity of single cell i and single cell j : :

[0034] (1)

[0035] Where is the temperature of single cell , is the temperature of single cell , is the state of charge of single cell , is the state of charge of single cell , is the initial capacity of single cell , is the initial capacity of single cell , is a constant to prevent the denominator from being zero, is a weight influence factor, and can be adjusted according to the actual situation.

[0036] In an embodiment of the present disclosure, based on historical sequential operation data, the health states of every two single cells in a battery pack can be obtained. According to the state of charge and health state of every two single cells, the health state weights of every two single cells can be determined, the distribution calibration terms of the health characteristic features of every two single cells can be obtained, and a health characteristic map can be constructed based on the health state weights, initial capacities, and distribution calibration terms of every two single cells.

[0037] Among them, the distribution calibration term, that is, the distribution calibration term of the health characteristic feature, is used to smooth the influence of outliers or noise.

[0038] Optionally, based on the health state weights, initial capacities, and distribution calibration terms of every two single cells, each element in the health characteristic matrix of the battery pack can be determined. Based on each element in the health characteristic matrix of the battery pack, a health characteristic matrix can be constructed, and the health characteristic matrix can be converted into a health characteristic map. 。

[0039] It should be noted that traditional health characteristic maps are constructed only based on the similarity of health indicators, such as the state of charge (SOC) and state of health (SOH). However, during the actual operation of the battery pack, the importance of health indicators may change dynamically.

[0040] For example, for a single cell in a battery pack i and a single cell j , based on the health state weights, initial capacities, and distribution calibration terms of single cell i and single cell j , the element in the health characteristic matrix of the battery pack can be determined :

[0041] (2)

[0042] Among them, is the health state weight, is the health state of single cell , is the health state of single cell , and are adjustable parameters, is the distribution calibration term.

[0043] For example, based on the distributions of SOC and SOH in historical sequential operation data, the health state weight can be adaptively updated according to the following formula :

[0044] (3)

[0045] Among them, is the single cell 's state of charge, is the single cell 's state of charge, is the single cell 's state of health, is the single cell 's state of health, is a constant to prevent the denominator from being zero.

[0046] In the embodiments of the present disclosure, the number of pre-divided time windows can be obtained, the correlation between every two single cells in the battery pack within each same time window can be obtained, the time decay parameter can be obtained, and based on the correlation between every two single cells in each same time window and the time decay parameter, a historical dependence graph of the battery pack can be constructed.

[0047] It should be noted that traditional historical dependence graphs often rely on simple time series correlations and are difficult to capture multi-scale characteristics over long time periods.

[0048] Optionally, each element in the historical dependence matrix of the battery pack can be determined based on the correlation between every two single cells in each same time window and the time decay parameter. Based on each element in the historical dependence matrix of the battery pack, a historical dependence matrix can be constructed, and the historical dependence matrix can be converted into a historical dependence graph .

[0049] For example, for the single cell i and the single cell j in the battery pack, the element i in the historical dependence matrix of the battery pack can be determined based on the correlation between the single cell j and the single cell within each same time window and the time decay parameter:

[0050] (4)

[0051] Wherein, is the number of time windows, m is the current time window, is the correlation between the single cell i and the single cell j within the time window m, is the feature matrix of the single cell i within the time window m, is the feature matrix of the single cell j within the time window m, is the time decay parameter.

[0052] S103. Input the feature matrix and the static adjacency graph into the dynamic graph generation network of the health state prediction model to obtain the dynamic adjacency graph output by the dynamic graph generation network.

[0053] Among them, the health state prediction model is pre-trained to predict the health state of the single cells in the battery pack. The health state prediction model includes a trained dynamic graph generation network (dynamic graph generator).

[0054] In the embodiment of the present disclosure, after obtaining the static adjacency graph, the feature matrix and the static adjacency graph can be input into the dynamic graph generation network of the health state prediction model to obtain the dynamic adjacency graph output by the dynamic graph generation network. A dyn 。

[0055] Among them, the dynamic adjacency graph can reflect the dynamic interaction relationship between the single cells in the battery pack over time.

[0056] S104. Input the multi-feature graph into the graph convolutional network of the health state prediction model to obtain the spatial features output by the graph convolutional network. Among them, the multi-feature graph includes the static adjacency graph, the health characteristic graph, the historical dependence graph, and the dynamic adjacency graph.

[0057] In the embodiment of the present disclosure, after obtaining the dynamic adjacency graph, the multi-feature graph can be determined based on the static adjacency graph, the health characteristic graph, the historical dependence graph, and the dynamic adjacency graph.

[0058] Among them, the health state prediction model further includes a trained graph convolutional network (Graph Convolutional Network, abbreviated as GCN).

[0059] In the embodiment of the present disclosure, after obtaining the multi-feature graph, the multi-feature graph can be input into the graph convolutional network of the health state prediction model, and convolution operations are performed on the multi-feature graph based on GCN to obtain the spatial features output by the graph convolutional network.

[0060] S105. Input the spatial features into the KAN network of the health state prediction model to obtain the health state prediction values of each single cell in the battery pack output by the KAN network.

[0061] Among them, the health state prediction model further includes a trained Kolmogorov–Arnold Network (abbreviated as KAN).

[0062] It should be noted that during the prediction of the state of health of the battery pack, the operating state of the battery pack is affected by the complex non-linear effects of multiple factors, such as temperature, charge and discharge rate, and the interaction between individual cells. To address the limitations of existing models in modeling high-dimensional non-linear characteristics, the present disclosure introduces KAN. By leveraging the powerful high-dimensional feature decomposition and reconstruction capabilities of KAN, the dynamic non-linear characteristics of the battery pack can be accurately characterized.

[0063] Among them, the Kolmogorov–Arnold theorem states that any continuous multi-dimensional non-linear function can be approximated by a combination of a series of univariate non-linear functions. This property gives KAN a significant advantage in dealing with complex high-dimensional non-linear data.

[0064] In the embodiments of the present disclosure, after obtaining the spatial features, the spatial features can be input into the KAN network of the state of health prediction model, and the spatial features are decomposed and reconstructed based on the KAN network to obtain the predicted values of the state of health of each individual cell in the battery pack output by the KAN network.

[0065] It should be noted that in a multi-battery system, due to the complex dynamic interaction and performance differentiation between individual cells, there are many challenges in predicting the state of health (SOH) value: (1) The operating states of individual cells in the battery pack are highly non-linear and dynamic, and their performance is affected by multiple factors such as temperature, current, and voltage. Traditional prediction models are difficult to accurately describe the complex operating mechanism of the battery pack; (2) There are complex spatial interaction relationships between individual cells in the battery pack. Ignoring these relationships will lead to insufficient prediction accuracy of the overall state of health of the battery pack. In addition, the actual operating data of the multi-battery system is often accompanied by noise and uncertainty, further increasing the difficulty of SOH prediction.

[0066] It should be noted that traditional SOH prediction methods mostly rely on static or predefined battery relationship graphs and are difficult to capture the dynamic interaction relationships that change over time in the battery pack. The present invention introduces a dynamic graph generation network that can obtain a dynamic adjacency graph in real time according to the dynamic changes of the operating characteristics of the battery pack, dynamically capture the complex interaction relationships between individual cells, thereby enhancing the model's adaptability to time-varying characteristics, fully integrating time and space dimension information, enhancing the modeling ability for the complex behaviors of the multi-battery system, extracting feature data including SOC range, charging capacity, voltage range, etc. for the characteristics of the multi-battery system, and constructing highly robust input features to improve the prediction accuracy of the state of health values of each individual cell in the battery pack.

[0067] In summary, the method for predicting the health state of a battery pack provided by the embodiments of the present disclosure obtains the historical time-series operation data of each single battery in the battery pack to be predicted, extracts features from the historical time-series operation data to obtain the feature matrix of the battery pack, constructs a static adjacency graph, a health characteristic graph, and a historical dependence graph of the battery pack based on the historical time-series operation data, inputs the feature matrix and the static adjacency graph into the dynamic graph generation network of the health state prediction model to obtain the dynamic adjacency graph output by the dynamic graph generation network, inputs the multi-feature graph into the graph convolutional network of the health state prediction model to obtain the spatial features output by the graph convolutional network, where the multi-feature graph includes a static adjacency graph, a health characteristic graph, a historical dependence graph, and a dynamic adjacency graph, and inputs the spatial features into the KAN network of the health state prediction model to obtain the health state prediction values of each single battery in the battery pack output by the KAN network. Thus, based on the trained health state prediction model, the present disclosure can effectively handle the high-dimensional non-linear complexity during the operation of the battery pack, accurately and efficiently predict the health state values of each single battery in the battery pack, and is beneficial to improving the operation safety and stability of the battery system.

[0068] The training process of the health state prediction model provided by the present disclosure will be explained below.

[0069] As Figure 2 shown, the training process of the health state prediction model proposed in this embodiment includes the following steps:

[0070] S201, obtain the sample time-series operation data of each sample single battery in the sample battery pack, and determine the true health state value of each sample single battery in the sample battery pack according to the sample time-series operation data.

[0071] In the embodiments of the present disclosure, after obtaining the sample time-series operation data, the health state value can be calculated based on the sample time-series operation data to determine the true health state value of each sample single battery in the sample battery pack.

[0072] S202, construct a training sample based on the sample time-series operation data and the true health state value.

[0073] In the embodiments of the present disclosure, after obtaining the sample time-series operation data and the true health state value, a training sample can be constructed based on the sample time-series operation data and the true health state value.

[0074] S203, train the initial health state prediction model to be trained according to the training sample to obtain the health state prediction model, where the health state prediction model includes a dynamic graph generation network, a graph convolutional network, and a KAN network.

[0075] In the embodiments of the present disclosure, the predicted health state values of each sample single battery in the sample battery pack output by the initial health state prediction model are obtained. Based on the predicted health state values of each sample single battery in the sample battery pack, the predicted correlation strength values of every two sample single batteries in the sample battery pack are obtained. And based on the true health state values of each sample single battery in the sample battery pack, the true correlation strength values of every two sample single batteries in the sample battery pack are obtained. According to the predicted health state values and the true health state values, a first loss function is determined, and a first weight of the first loss function is obtained. According to the predicted correlation strength values and the true correlation strength values, a second loss function is determined, and a second weight of the second loss function is obtained. A regularization term is obtained, and a third weight of the regularization term is obtained. Based on the first loss function and the first weight, the second loss function and the second weight, and the regularization term and the third weight, the total loss function of the health state prediction model is determined. Based on the total loss function, the model parameters of the initial health state prediction model are adjusted, and the adjusted health state prediction model is continuously trained until the training is completed, thereby obtaining the health state prediction model.

[0076] In the embodiments of the present disclosure, the expression of the total loss function may be:

[0077] (5)

[0078] Wherein, is the total loss function, is the first loss function, is the first weight, is the second loss function, is the second weight, is the regularization term, is the third weight.

[0079] It should be noted that the first loss function can be used to describe the loss of time dynamic characteristics. By calculating the deviation between the predicted health state value and the true health state value step by step in time, the first loss function can effectively capture the ability of the health state prediction model to model the dynamic changes of the battery pack state in the time dimension. The smaller the value of the first loss function, the better the health state prediction model can track the time evolution of the battery pack health state.

[0080] Optionally, the expression of the first loss function is:

[0081] (6)

[0082] Wherein, is the number of single batteries in the battery pack, is the number of time steps, is the predicted value of the health state of the single battery i at time t, is the true value of the health state of the single battery i at time t, and p is 1 or 2.

[0083] Optionally, a second loss function can be used to describe the loss of spatial dynamic characteristics. The expression of the second loss function is:

[0084] (7)

[0085] (8)

[0086] where is the predicted value of the correlation strength between the battery cell i and the single battery cell j at time t, is the true value of the correlation strength between the battery cell i and the single battery cell j at time t, is 1 or 2.

[0087] Optionally, the correlation strength value of every two sample single battery cells can be calculated based on the Pearson correlation coefficient.

[0088] In the embodiment of the present disclosure, the model parameters of the health state prediction model can be continuously adjusted according to the total loss function until the loss function meets the training end condition, and the health state prediction model after the last adjustment of the model parameters is determined as the trained health state prediction model.

[0089] It should be noted that the setting of the training end condition in this application is not limited and can be selected according to the actual situation. Optionally, the training end condition can be set that the value of the total loss function is less than a preset loss threshold.

[0090] In summary, the battery pack health state prediction method provided by the embodiments of the present disclosure obtains the sample time-series operation data of each sample single battery in the sample battery pack, determines the true value of the health state of each sample single battery in the sample battery pack according to the sample time-series operation data, constructs training samples based on the sample time-series operation data and the true value of the health state, and trains the initial health state prediction model to be trained according to the training samples to obtain a health state prediction model. Among them, the health state prediction model includes a dynamic graph generation network, a graph convolutional network, and a KAN network. Thus, the present disclosure combines the collaborative modeling capabilities of the dynamic graph generation network, the graph convolutional network, and the KAN network. The dynamic graph generation network is used to capture the dynamic relationships that change over time, the graph convolutional network extracts spatial features, and the KAN models the time dynamics and non-linear characteristics, realizing the joint modeling of time dynamic characteristics and the spatial dependence of multiple single batteries. It can dynamically capture the complex interaction relationships between single batteries, significantly enhance the model's expression ability for the spatio-temporal interaction characteristics of the battery pack, comprehensively describe the interaction dependence relationships between multiple single batteries, have stronger adaptability and generalization ability, be suitable for the prediction requirements of complex systems such as energy storage power stations, achieve high-precision prediction of the health state, and provide technical support for the intelligent management and operation and maintenance of the battery pack.

[0091] As a possible implementation, as Figure 3 shown, on the basis of the above embodiment, the specific process of the dynamic adjacency graph output by the dynamic graph generation network includes the following steps:

[0092] S301, based on the feature matrix, obtain the dynamic embedding vector of each single battery in the battery pack.

[0093] In the embodiments of the present disclosure, the feature vector of the single battery in the battery pack can be obtained based on the feature matrix, and the dynamic embedding vector of each single battery in the battery pack can be obtained based on the feature vector.

[0094] For example, for the single battery i at time t, the feature vector , fuse the feature matrix i of the single battery at time t, the time information vector and the hidden state i of the single battery at the previous time t - 1 to provide time series dependence information, that is .

[0095] For example, for the dynamic embedding vector of the single battery i in the battery pack at time t, it can be obtained according to the following formula:

[0096] (9)

[0097] Among them, is the dynamic embedding vector of the single cell at time t, i is the dynamic embedding vector of the single cell at time t, is the single cell i is the feature vector of the single cell at time t, is the activation function, is the weight matrix, is the bias vector.

[0098] S302. Obtain the dynamic similarity between every two single cells according to the dynamic embedding vector of each single cell, and obtain the first weight of the dynamic similarity.

[0099] It should be noted that the traditional Gaussian kernel function usually assumes that the characteristic changes between single cells are smooth, and ignores the non-synchronization and multi-scale dependence problems of time series. The present disclosure constructs a similarity metric framework combining multi-scale Gaussian kernel and dynamic time warping (DTW) to accurately capture the multi-dimensional characteristics of the dynamic interaction between single cells in a battery pack.

[0100] For example, for the single cells i and the single cell j in the battery pack, the dynamic similarity can be obtained according to the following formula:

[0101] (10)

[0102] Among them, is the dynamic similarity between the single cell i and the single cell j at time t, is the dynamic embedding vector of the single cell i at time t, is the dynamic embedding vector of the single cell j at time t, is the similarity metric calculated by DTW, is the number of scales of the Gaussian kernel function, is the k th scale weight, satisfying , used to balance the contributions of different scales, the k th scale standard deviation, used to adjust the kernel function width of different scales.

[0103] S303. Obtain the second weight of the static adjacency graph.

[0104] S304. Obtain the dynamic influence function and the third weight of the dynamic influence function.

[0105] It should be noted that the dynamic influence function can represent the time decay weights of features such as current and temperature.

[0106] Optionally, the expression of the dynamic influence function can be:

[0107] (11)

[0108] Where is the time decay coefficient, ∣ t i -t j ∣ is the time difference between the single cell i and the single cell j .

[0109] It should be noted that by adding the time decay coefficient, the dynamic characteristics of the interaction relationship between single cells changing with time can be described more flexibly, which is especially suitable for capturing long-range dependencies.

[0110] S305. Based on the first weight and the dynamic similarity, the second weight and the static adjacency graph, and the third weight and the dynamic influence function, obtain the dynamic adjacency graph.

[0111] It should be noted that the first weight , the second weight and the third weight are adjustable weight parameters, which can be used to control the influence degree of the dynamic interaction relationship and the static interaction relationship.

[0112] Optionally, each element in the dynamic adjacency matrix of the battery pack can be determined based on the first weight and the dynamic similarity, the second weight and the static adjacency graph, and the third weight and the dynamic influence function. Based on each element in the dynamic adjacency matrix of the battery pack, where each element can be used to describe the dynamic interaction relationship between every two single cells.

[0113] Construct a dynamic adjacency matrix and convert the dynamic adjacency matrix into a dynamic adjacency graph A dyn .

[0114] For example, for the single cell i and the single cell j in the battery pack, the element in the dynamic adjacency matrix of the battery pack can be determined based on the first weight and the dynamic similarity, the second weight and the static adjacency graph, and the third weight and the dynamic influence function:

[0115] (12)

[0116] Where is the single celli and single cells j The dynamic interaction relationship at time t, is the first weight, for the single cell i and single cells j The dynamic similarity at time t, is an element in the static adjacency matrix, is the second weight, for the single cell i and single cells j The dynamic influence function value at time t, is the third weight, is the outer weight parameter, which determines the comprehensive influence of the dynamic similarity, static adjacency graph and dynamic influence function.

[0117] In the embodiments of the present disclosure, after obtaining the dynamic adjacency graph, the multi-feature graph can be input into the GCN of the health state prediction model. The multi-feature graph not only covers the static and dynamic interaction relationships between single cells, but also extends to health state, historical association and environmental dependence. By aggregating the neighborhood information of the multi-feature graph through GCN, complex spatial features are extracted.

[0118] It should be noted that by obtaining the dynamic adjacency graph, the structure of the multi-feature graph can be adjusted according to the change of time, so as to more flexibly adapt to the dynamic change of the operating conditions of the battery pack, making the performance of GCN in spatial feature modeling significantly better than the traditional method based on static graph, providing higher accuracy and applicability for health state prediction.

[0119] As a possible implementation, as Figure 4 shown, on the basis of the above embodiments, the specific process of the spatial features output by the graph convolutional network includes the following steps:

[0120] S401, Obtain the adjacency matrix and degree matrix of each graph in the multi-feature graph, and normalize the adjacency matrix of each graph according to the degree matrix to obtain the normalized adjacency matrix of each graph.

[0121] In the embodiments of the present disclosure, for the k-th feature graph in the multi-feature graph, the normalized adjacency matrix of the k-th feature graph can be expressed as , is the degree matrix of the k-th feature graph, is the adjacency matrix of the k-th feature graph.

[0122] S402, Perform graph convolution on the normalized adjacency matrix of each graph to obtain aggregated features.

[0123] In the embodiments of the present disclosure, for each graph, convolution is performed based on the normalized adjacency matrix and the feature matrix, and the convolution results of all graphs are added together to obtain the aggregated features:

[0124] (13)

[0125] Among them, is the number of multi-feature graphs, is the spatial feature of the layer, and

[0126] is the weight matrix of the k-th feature graph.

[0127] Optionally, the spatial features can be obtained according to the following formula :

[0128] (14)

[0129] Among them, is the spatial feature of the layer, is the aggregated feature, is the regularization bias term, is the hyperparameter,

[0130] As a possible implementation, as Figure 5 shown, on the basis of the above embodiments, the specific process of the health state prediction value of each single battery in the battery pack output by the KAN network includes the following steps:

[0131] S501, perform linear transformation and non-linear activation on the spatial features to obtain the decomposed features.

[0132] In the embodiments of the present disclosure, after obtaining the spatial features, other features with a time series relationship are obtained, and the spatial features and other features with a time series relationship are summarized to obtain the target feature matrix including all single batteries in the battery pack, where the target feature matrix contains features.

[0133] In the embodiments of the present disclosure, after obtaining the spatial features, the projection of the spatial features in the time dimension is obtained to obtain the time series features , perform linear transformation and non-linear activation on the spatial features to obtain the decomposed features:

[0134] (15)

[0135] Among them, is the decomposed feature, is the activation function, and are the weight parameters of feature decomposition, is the target feature matrix.

[0136] S502, based on the non-uniform B-spline kernel function, perform non-linear transformation and reconstruction on the decomposed features to obtain the predicted value of the health state of each single battery in the battery pack.

[0137] In the embodiments of the present disclosure, the predicted value of the health state of each single battery in the battery pack can be obtained according to the following formula:

[0138] (16)

[0139] Among them, is the health state prediction function, is the feature in the target feature matrix (high-dimensional feature), is the univariate non-linear transformation, is the projection function of the high-dimensional feature, Q is the decomposition dimension in KAN.

[0140] It should be noted that, compared with traditional deep learning methods, KAN allows the health state prediction model to dynamically adjust the weights of different non-linear transformations in feature fusion, improving the ability to capture complex dynamic characteristics. Through the feature decomposition and reconstruction mechanism, the problem of gradient disappearance commonly encountered in complex time series modeling is avoided. KAN can not only effectively reduce the dimension of input features, convert high-dimensional features into a combination of multiple univariate variables, significantly reducing the computational complexity of the model, but also fully retain the dynamic change information of high-dimensional features, greatly improving the accuracy of health state prediction. At the same time, the non-linear expression ability of KAN enables the comprehensive characterization of the complex dynamic characteristics of the health state of the battery pack, especially suitable for the complex operating environment of multi-module systems such as energy storage power stations.

[0141] It should be noted that, to further improve the non-linear representation ability, the present disclosure combines a non-uniform B-spline kernel function in KAN to perform non-linear transformation and reconstruction on the decomposed features. By adjusting the characteristics of the B-spline kernel function through dynamic node distribution, the non-uniform changes of input features are captured. The dynamic node distribution is generated by the following formula:

[0142] (17)

[0143] Among them, is the dynamic distribution value of the k-th node, is the characteristic difference value between nodes, and N is the total number of spline nodes.

[0144] It should be noted that, based on the dynamic node distribution, the expression of the non-uniform B-spline basis function is:

[0145] (18)

[0146] Among them, is the decomposed feature is the weight of the basis function, is the specific form of the B-spline basis function, which controls the smoothness and locality of the non-linear mapping, is the decomposed feature, is the dynamic distribution value of the k-th node.

[0147] In the embodiment of the present disclosure, after obtaining the features after non-linear transformation, can be reconstructed according to the following formula:

[0148] (19)

[0149] Among them, is the weight of the feature reconstruction, is the weight of the non-uniform B-spline basis function.

[0150] In summary, the battery pack health state prediction method provided by the embodiment of the present disclosure can obtain a dynamic adjacency graph through dynamic similarity (temporal dynamic characteristic similarity), static adjacency graph, and dynamic influence function, which can reflect the dynamic changes of the interaction relationship during the operation of the battery pack, solve the limitation that the traditional static graph modeling method is difficult to adapt to time-varying systems, realize the joint modeling of temporal dynamic characteristics and multi-module spatial dependence relationships by combining GCN and KAN, and can capture the complex dynamic characteristics in the change of battery health state through high-dimensional non-linear decomposition and reconstruction of spatial features by KAN, significantly improving the expression ability of non-linear and time-dependent characteristics, and significantly improving the accuracy of the predicted value of the health state of each single battery in the battery pack.

[0151] The following explains the specific process of the battery pack health state prediction method proposed by the present disclosure.

[0152] For example, such as Figure 6As shown, obtain the historical sequential operation data of each single battery in the battery pack, extract features from the historical sequential operation data to obtain the feature matrix of the battery pack, construct the static adjacency graph, health characteristic graph, and historical dependence graph of the battery pack based on the historical sequential operation data, input the feature matrix and the static adjacency graph into the dynamic graph generation network of the health state prediction model to obtain the dynamic adjacency graph output by the dynamic graph generation network, input the multi-feature graphs (static adjacency graph, health characteristic graph, historical dependence graph, and dynamic adjacency graph) into the graph convolutional network of the health state prediction model to construct spatial features, so as to obtain the spatial features output by the graph convolutional network, and input the spatial features into the KAN network of the health state prediction model for feature decomposition and reconstruction to obtain the health state prediction values of each single battery in the battery pack output by the KAN network.

[0153] In summary, based on the trained health state prediction model, the present disclosure improves the prediction accuracy and efficiency of predicting the health state values of each single battery in the battery pack, enhances the robustness and stability of the health state prediction model under different operating conditions, can be applied to various complex working conditions and operating conditions in practical application scenarios, can effectively improve the operating safety and stability of the battery system, can be applied to lithium-ion battery packs or other energy storage devices or power battery management systems, etc., and provides important technical support for the health management in the fields of electric energy storage, electric transportation, renewable energy, etc.

[0154] Figure 7 is a block diagram of a device for predicting the health state of a battery pack shown according to an exemplary embodiment. As Figure 7 shown, the device 700 for predicting the health state of a battery pack according to an embodiment of the present disclosure may specifically include: a first acquisition module 701, a second acquisition module 702, a first processing module 703, a first processing module 703, and a third processing module 705.

[0155] The first acquisition module 701 is configured to acquire the historical sequential operation data of each single battery in the battery pack to be predicted, extract features from the historical sequential operation data, and obtain the feature matrix of the battery pack;

[0156] The second acquisition module 702 is configured to construct the static adjacency graph, health characteristic graph, and historical dependence graph of the battery pack based on the historical sequential operation data;

[0157] The first processing module 703 is configured to input the feature matrix and the static adjacency graph into the dynamic graph generation network of the health state prediction model to obtain the dynamic adjacency graph output by the dynamic graph generation network;

[0158] A second processing module 704, configured to input multi-feature maps into a graph convolutional network of the health state prediction model to obtain spatial features output by the graph convolutional network, where the multi-feature maps include the static adjacency graph, the health characteristic graph, the historical dependency graph, and the dynamic adjacency graph;

[0159] A third processing module 705, configured to input the spatial features into a KAN network of the health state prediction model to obtain health state prediction values of each single battery in the battery pack output by the KAN network.

[0160] In an embodiment of the present disclosure, the second processing module 704 is further configured to: based on the historical time-series operation data, obtain the temperature, state of charge, and initial capacity of every two single batteries in the battery pack; based on the temperature, state of charge, and initial capacity of every two single batteries, construct the static adjacency graph of the battery pack.

[0161] In an embodiment of the present disclosure, the second processing module 704 is further configured to: based on the historical time-series operation data, obtain the health state of every two single batteries in the battery pack; determine the health state weight of every two single batteries according to the state of charge and health state of every two single batteries; obtain the distribution calibration term of the health characteristic features of every two single batteries; based on the health state weight, the initial capacity, and the distribution calibration term of every two single batteries, construct the health characteristic graph.

[0162] In an embodiment of the present disclosure, the second processing module 704 is further configured to: obtain the number of pre-divided time windows, and obtain the correlation between every two single batteries in the battery pack within each same time window; obtain a time decay parameter, and construct the historical dependency graph of the battery pack according to the correlation between every two single batteries within each same time window and the time decay parameter.

[0163] In an embodiment of the present disclosure, the training process of the health state prediction model includes: obtaining the sample time-series operation data of each sample single battery in the sample battery pack, and determining the true value of the health state of each sample single battery in the sample battery pack according to the sample time-series operation data; constructing a training sample based on the sample time-series operation data and the true value of the health state; training an initial health state prediction model to be trained according to the training sample to obtain the health state prediction model, where the health state prediction model includes a dynamic graph generation network, a graph convolutional network, and a KAN network.

[0164] In one embodiment of the present disclosure, the training process of the health state prediction model includes: obtaining the health state prediction value of each sample single cell in the sample battery pack output by the initial health state prediction model;

[0165] Based on the health state prediction values of each sample single cell in the sample battery pack, obtaining the correlation strength prediction values of every two sample single cells in the sample battery pack, and based on the true health state values of each sample single cell in the sample battery pack, obtaining the true correlation strength values of every two sample single cells in the sample battery pack;

[0166] According to the health state prediction value and the true health state value, determining a first loss function and obtaining a first weight of the first loss function; according to the correlation strength prediction value and the true correlation strength value, determining a second loss function and obtaining a second weight of the second loss function; obtaining a regularization term and obtaining a third weight of the regularization term; based on the first loss function and the first weight, the second loss function and the second weight, the regularization term and the third weight, determining the total loss function of the health state prediction model; based on the total loss function, adjusting the model parameters of the initial health state prediction model and continuing to train the adjusted health state prediction model until the training ends to obtain the health state prediction model.

[0167] In one embodiment of the present disclosure, the first processing module 703 is configured to: based on the feature matrix, obtain the dynamic embedding vector of each single cell in the battery pack; according to the dynamic embedding vector of each single cell, obtain the dynamic similarity of every two single cells and obtain a first weight of the dynamic similarity; obtain a second weight of the static adjacency graph; obtain a dynamic influence function and a third weight of the dynamic influence function; based on the first weight and the dynamic similarity, the second weight and the static adjacency graph, the third weight and the dynamic influence function, to obtain the dynamic adjacency graph.

[0168] In one embodiment of the present disclosure, the second processing module 704 is configured to: obtain the adjacency matrix and degree matrix of each graph in the multi-feature graph, and normalize the adjacency matrix of each graph according to the degree matrix to obtain the normalized adjacency matrix of each graph; perform graph convolution on the normalized adjacency matrix of each graph to obtain aggregated features; process the aggregated features according to an activation function, and based on the processed features and a regularization bias term, to obtain the spatial features.

[0169] In one embodiment of the present disclosure, the third processing module 705 is configured to: perform a linear transformation and a non-linear activation on the spatial features to obtain decomposed features; and perform a non-linear transformation and reconstruction on the decomposed features based on a non-uniform B-spline kernel function to obtain the predicted health state values of each single battery in the battery pack.

[0170] In the embodiments of the present disclosure, the specific manners in which the respective modules in the battery pack health state prediction device of the above embodiments perform operations have been described in detail in the embodiments related to the battery pack health state prediction method, and will not be elaborated herein.

[0171] In summary, the battery pack health state prediction device provided by the embodiments of the present disclosure obtains the historical sequential operation data of each single battery in the battery pack to be predicted, extracts features from the historical sequential operation data to obtain a feature matrix of the battery pack, constructs a static adjacency graph, a health characteristic graph, and a historical dependency graph of the battery pack based on the historical sequential operation data, inputs the feature matrix and the static adjacency graph into the dynamic graph generation network of the health state prediction model to obtain a dynamic adjacency graph output by the dynamic graph generation network, inputs the multi-feature graph into the graph convolutional network of the health state prediction model to obtain spatial features output by the graph convolutional network, where the multi-feature graph includes the static adjacency graph, the health characteristic graph, the historical dependency graph, and the dynamic adjacency graph, and inputs the spatial features into the KAN network of the health state prediction model to obtain the predicted health state values of each single battery in the battery pack output by the KAN network. Thus, the present disclosure can effectively cope with the high-dimensional non-linear complexity of the operating characteristics of the battery pack, and based on the trained health state prediction model, can accurately and efficiently predict the health state values of each single battery in the battery pack, which is beneficial to improving the operating safety and stability of the battery system.

[0172] Figure 8 It is a block diagram of an electronic device 1000 shown according to an exemplary embodiment.

[0173] As Figure 8 shown, the above electronic device 1000 includes:

[0174] A memory 1001 and a processor 1002, a bus 1003 connecting different components (including the memory 1001 and the processor 1002), and the memory 1001 stores a computer program, and when the processor 1002 executes the program, the battery pack health state prediction method of the embodiments of the present disclosure is implemented.

[0175] Bus 1003 represents one or more of several types of bus architectures, including a memory bus or memory controller, a peripheral bus, an Accelerated Graphics Port, a processor bus, or a local bus using any of the several bus architectures. By way of example, and not limitation, such architectures include the Industry Standard Architecture (ISA) bus, Micro Channel Architecture (MAC) bus, Enhanced ISA bus, Video Electronics Standards Association (VESA) local bus, and Peripheral Component Interconnect (PCI) bus.

[0176] Electronic device 1000 typically includes a variety of electronic device readable media. These media can be any available media that can be accessed by electronic device 1000, including both volatile and nonvolatile media, removable and non-removable media.

[0177] Memory 1001 can also include computer system readable media in the form of volatile memory, such as random access memory (RAM) 1004 and / or cache memory 1005. Electronic device 1000 can further include other removable / non-removable, volatile / nonvolatile computer system storage media. By way of example only, storage system 1006 can be used for reading and writing non-removable, nonvolatile magnetic media ( Figure 8 not shown and typically called a "hard disk drive"). Although Figure 8 not shown in the figures, a disk drive for reading and writing removable nonvolatile disks (such as a "floppy disk"), and an optical disk drive for reading and writing removable nonvolatile optical disks (such as a CD-ROM, DVD-ROM or other optical media) can be provided. In these cases, each drive can be connected to bus 1003 by one or more data media interfaces. Memory 1001 can include at least one program product having a set (e.g., at least one) of program modules that are configured to carry out the functions of the embodiments of the present disclosure.

[0178] A program / utility 1008 having a set (at least one) of program modules 1007 can be stored in, for example, memory 1001, such program modules 1007 including but not limited to an operating system, one or more application programs, other program modules, and program data, each of which examples or some combination thereof may include an implementation of a network environment. Program modules 1007 typically carry out the functions and / or methods of the embodiments described herein.

[0179] The electronic device 1000 can also communicate with one or more external devices 1009 (such as a keyboard, a pointing device, a display 1011, etc.), and can also communicate with one or more devices that enable a user to interact with the electronic device 1000, and / or communicate with any device that enables the electronic device 1000 to communicate with one or more other computing devices (such as a network card, a modem, etc.). Such communication can be carried out through the input / output (I / O) interface 1012. Moreover, the electronic device 1000 can also communicate with one or more networks (such as a local area network (LAN), a wide area network (WAN), and / or a public network, such as the Internet) through the network adapter 1013. As Figure 8 shown, the network adapter 1013 communicates with other modules of the electronic device 1000 through the bus 1003. It should be understood that, although not shown in the figure, other hardware and / or software modules can be used in combination with the electronic device 1000, including but not limited to: microcode, device drivers, redundant processing units, external disk drive arrays, RAID systems, tape drives, and data backup storage systems, etc.

[0180] The processor 1002 executes various functional applications and data processing by running programs stored in the memory 1001.

[0181] It should be noted that for the implementation process and technical principle of the electronic device in this embodiment, refer to the foregoing explanation of the battery pack health state prediction method of the embodiments of the present disclosure, and details are not described herein again.

[0182] To implement the above embodiments, the present disclosure also proposes a computer-readable storage medium.

[0183] Wherein, when the instructions in the computer-readable storage medium are executed by the processor of the electronic device, the electronic device can execute the battery pack health state prediction method as described above. Optionally, the computer-readable storage medium can be a ROM, a random access memory (RAM), a CD-ROM, a magnetic tape, a floppy disk, and an optical data storage device, etc.

[0184] Those skilled in the art will readily think of other implementation schemes of the present disclosure after considering the specification and practicing the invention disclosed herein. The present disclosure aims to cover any variations, uses, or adaptive changes of the present disclosure, and these variations, uses, or adaptive changes follow the general principles of the present disclosure and include common general knowledge or conventional technical means in the technical field not disclosed in the present disclosure.

[0185] It should be understood that the present disclosure is not limited to the exact structures described above and shown in the drawings, and various modifications and changes can be made without departing from its scope.

Claims

1. A method for predicting the state of health of a battery pack, characterized in that, The method includes: Obtaining historical sequential operation data of each single battery in the battery pack to be predicted, extracting features from the historical sequential operation data, and obtaining a feature matrix of the battery pack; Based on the historical sequential operation data, constructing a static adjacency graph, a health characteristic graph, and a historical dependence graph of the battery pack; Inputting the feature matrix and the static adjacency graph into a dynamic graph generation network of a health state prediction model to obtain a dynamic adjacency graph output by the dynamic graph generation network; Inputting a multi-feature graph into a graph convolutional network of the health state prediction model to obtain spatial features output by the graph convolutional network, where the multi-feature graph includes the static adjacency graph, the health characteristic graph, the historical dependence graph, and the dynamic adjacency graph; Inputting the spatial features into a KAN network of the health state prediction model to obtain health state prediction values of each single battery in the battery pack output by the KAN network.

2. The method according to claim 1, wherein The constructing the static adjacency graph of the battery pack based on the historical sequential operation data includes: Based on the historical sequential operation data, obtaining the temperature, state of charge, and initial capacity of every two single batteries in the battery pack; Based on the temperature, state of charge, and initial capacity of every two single batteries, constructing the static adjacency graph of the battery pack.

3. The method according to claim 2, wherein The constructing the health characteristic graph of the battery pack based on the historical sequential operation data includes: Based on the historical sequential operation data, obtaining the health state of every two single batteries in the battery pack; According to the state of charge and health state of every two single batteries, determining the health state weight of every two single batteries; Obtaining a distribution calibration term of the health characteristic features of every two single batteries; Based on the health state weight, the initial capacity, and the distribution calibration term of every two single batteries, constructing the health characteristic graph.

4. The method according to claim 1, wherein The constructing the historical dependence graph of the battery pack based on the historical sequential operation data includes: Obtaining the number of pre-divided time windows, and obtaining the correlation of every two single batteries in the battery pack within each same time window; Obtaining a time decay parameter, and constructing the historical dependence graph of the battery pack according to the correlation of every two single batteries within each same time window and the time decay parameter.

5. The method according to claim 1, wherein The training process of the health state prediction model includes: Obtaining sample sequential operation data of each sample single battery in a sample battery pack, and determining the true health state value of each sample single battery in the sample battery pack according to the sample sequential operation data; Based on the sample sequential operation data and the true health state value, constructing a training sample; Training an initial health state prediction model to be trained according to the training sample to obtain the health state prediction model, where the health state prediction model includes a dynamic graph generation network, a graph convolutional network, and a KAN network.

6. The method according to claim 5, wherein The training the initial health state prediction model to be trained according to the training sample to obtain the health state prediction model includes: Obtain the predicted health state values of each sample single cell in the sample battery pack output by the initial health state prediction model; Based on the predicted health state values of each sample single cell in the sample battery pack, obtain the predicted correlation strength values of every two sample single cells in the sample battery pack, and based on the true health state values of each sample single cell in the sample battery pack, obtain the true correlation strength values of every two sample single cells in the sample battery pack; Determine a first loss function according to the predicted health state values and the true health state values, and obtain a first weight of the first loss function; Determine a second loss function according to the predicted correlation strength values and the true correlation strength values, and obtain a second weight of the second loss function; Obtain a regularization term, and obtain a third weight of the regularization term; Based on the first loss function and the first weight, the second loss function and the second weight, the regularization term and the third weight, determine the total loss function of the health state prediction model; Based on the total loss function, adjust the model parameters of the initial health state prediction model, and continue to train the adjusted health state prediction model until the training ends to obtain the health state prediction model.

7. The method according to claim 1, wherein The process of the dynamic graph generation network outputting the dynamic adjacency graph includes: Based on the feature matrix, obtain the dynamic embedding vectors of each single cell in the battery pack; According to the dynamic embedding vectors of each single cell, obtain the dynamic similarity of every two single cells, and obtain a first weight of the dynamic similarity; Obtain a second weight of the static adjacency graph; Obtain a dynamic influence function and a third weight of the dynamic influence function; Based on the first weight and the dynamic similarity, the second weight and the static adjacency graph, the third weight and the dynamic influence function, to obtain the dynamic adjacency graph.

8. The method according to claim 1, characterized in that The process of the graph convolutional network outputting the spatial features includes: Obtain the adjacency matrix and degree matrix of each graph in the multi-feature graph, and normalize the adjacency matrix of each graph according to the degree matrix to obtain the normalized adjacency matrix of each graph; Perform graph convolution on the normalized adjacency matrix of each graph to obtain aggregated features; According to the activation function, process the aggregated features, and based on the processed features and the regularization bias term, to obtain the spatial features.

9. The method according to claim 1, characterized in that The process of the KAN network outputting the predicted health state values of each single cell in the battery pack includes: Perform a linear transformation and a non-linear activation on the spatial features to obtain decomposed features; Based on the non-uniform B-spline kernel function, perform a non-linear transformation and reconstruction on the decomposed features to obtain the predicted health state values of each single cell in the battery pack.

10. A battery pack state of health prediction device, characterized in that, The device includes: A first acquisition module, configured to acquire the historical time-series operation data of each single cell in the battery pack to be predicted, perform feature extraction on the historical time-series operation data, and acquire the feature matrix of the battery pack; A second acquisition module, configured to construct a static adjacency graph, a health characteristic graph, and a historical dependency graph of the battery pack based on the historical sequential operation data; A first processing module, configured to input the feature matrix and the static adjacency graph into a dynamic graph generation network of a health state prediction model to obtain a dynamic adjacency graph output by the dynamic graph generation network; A second processing module, configured to input a multi-feature graph into a graph convolutional network of the health state prediction model to obtain spatial features output by the graph convolutional network, where the multi-feature graph includes the static adjacency graph, the health characteristic graph, the historical dependency graph, and the dynamic adjacency graph; A third processing module, configured to input the spatial features into a KAN network of the health state prediction model to obtain health state prediction values of each single battery in the battery pack output by the KAN network.

Images