Battery state prediction method and device based on battery time sequence model

Abstract

The invention discloses a battery state prediction method and device based on a battery time sequence model, and belongs to the technical field of deep learning and battery management. The method comprises the following steps: acquiring a battery multivariate time sequence data set, and preprocessing the battery multivariate time sequence data set; constructing a battery time sequence model, wherein the battery time sequence model comprises a time-frequency feature extraction module, a multi-scale adaptive graph convolutional network and a multi-scale multi-frequency fusion module; the time-frequency feature extraction module extracts time-frequency mixed features in the battery multivariate time series data set, and the multi-scale adaptive graph convolutional network captures multi-scale features in the time-frequency mixed features and processes the multi-scale features to obtain a multi-scale tensor; the multi-scale multi-frequency fusion module fuses the multi-scale tensors on the frequency scale and the time scale to obtain fusion features; and predicting the state of the battery through the fused features. According to the method, the periodic mode and the non-periodic change characteristics under different time scales can be effectively captured, and the accuracy and the robustness of battery state prediction are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical fields of deep learning and battery management, and more specifically, to a battery state prediction method and device based on a battery time series model. Background Art

[0002] In recent years, new energy vehicles have become a new trend in the development of the automotive industry, with the market scale continuously expanding and technologies constantly innovating. The core component of a new energy vehicle is the power battery, and its performance directly affects the vehicle's cruising range, charge-discharge efficiency, service life, etc. While new energy vehicles are rapidly popularized, the prediction of power battery states has also become the focus of the industry. However, with the increase in the complexity of battery systems, there are still many deficiencies in existing technologies when dealing with the multi-dimensional characteristics and complex working conditions of batteries.

[0003] Currently, the existing technologies for power battery state prediction mainly include: (1) methods based on physical models, such as equivalent circuit models and electrochemical models based on thermodynamics. This technology requires a large amount of experimental data and a complex modeling process, is difficult to adapt to different battery types and dynamic working conditions, and is highly sensitive to environmental conditions, with the prediction accuracy decreasing in extreme cases such as high temperature or low temperature. (2) data-driven machine learning methods, such as support vector machines and neural networks. Although this method can capture complex relationships, it is insufficient in dealing with multi-scale correlations in battery data.

[0004] Power battery data usually has multi-scale characteristics, such as long-term battery aging trends and short-term current fluctuation patterns. However, existing methods are difficult to capture these multi-scale features simultaneously. The state parameters of power batteries have obvious periodic characteristics, and existing state prediction methods have significant deficiencies in discovering and utilizing these periodic characteristics. There is an important interaction relationship between the time dimension and frequency dimension of time series, but existing methods often process time or frequency features separately and are difficult to build the synergistic effect between the two.

[0005] Chinese Patent Application, Application No. 202411567374.1, Publication Date: December 27, 2024, discloses a method and system for predicting the state of charge of a battery based on a convolutional neural network. The method includes: obtaining the power consumption data of the battery; preprocessing the power consumption data; obtaining the first predicted SOC of the battery at each sampling time according to the dual Kalman filter algorithm and the power consumption data; forming a data set with the power consumption data and the first predicted SOC; dividing the data set into a training set, a validation set and a test set; constructing a convolutional neural network model CNN; training the convolutional neural network model with the training set to obtain a trained convolutional neural network model; validating the trained convolutional neural network model with the validation set to obtain a validated convolutional neural network model; testing the validated convolutional neural network model with the test set to obtain the second predicted SOC of the battery at the next moment. This method can only improve the quality of the battery data set to reduce the introduction of errors, but does not consider the multi-scale characteristics of the battery data. Therefore, there are significant deficiencies in discovering and utilizing the multi-scale characteristics of the battery data. Summary of the Invention

[0006] 1. Technical Problems to be Solved

[0007] Aiming at the problems of insufficient discovery of the periodic characteristics and multi-scale characteristics modeling of the battery state parameters in the prior art, the present invention provides a method and device for predicting the battery state based on a battery time series model. Through a multi-scale adaptive graph convolutional network, discrete Fourier transform and orthogonal basis functions, it can better capture the multi-scale characteristics of the battery and the periodic pattern characteristics at different time scales, effectively improving the accuracy and robustness of the battery state prediction.

[0008] 2. Technical Solutions

[0009] The object of the present invention is achieved through the following technical solutions.

[0010] A method for predicting the battery state based on a battery time series model includes the following steps:

[0011] Obtain a battery multivariate time series data set and preprocess the battery multivariate time series data set;

[0012] Construct a battery time series model, where the battery time series model includes a time-frequency feature extraction module, a multi-scale adaptive graph convolutional network and a multi-scale multi-frequency fusion module;

[0013] Use the time-frequency feature extraction module to extract the time-frequency hybrid features in the preprocessed battery multivariate time series data set; the multi-scale adaptive graph convolutional network captures the multi-scale features in the time-frequency hybrid features and processes the multi-scale features to obtain a multi-scale tensor; the multi-scale multi-frequency fusion module fuses the multi-scale tensor in the frequency scale and the time scale to obtain a fusion feature;

[0014] Predict the battery state through fused features.

[0015] As a further improvement of the present invention, the time-frequency feature extraction module extracts time-frequency hybrid features from the preprocessed battery multi-temporal data set, and the specific steps include:

[0016] Perform a discrete Fourier transform on the preprocessed battery multi-temporal data set to obtain a complex spectrum, and the complex spectrum includes a real component and an imaginary component;

[0017] Multiply the real component by an orthogonal cosine basis function to obtain a time series feature, multiply the imaginary component by an orthogonal sine basis function to obtain a frequency domain feature, and add the time series feature and the frequency domain feature to obtain a time-frequency hybrid feature.

[0018] As a further improvement of the present invention, the multi-scale adaptive graph convolutional network includes a multi-head attention mechanism, an adaptive adjacency matrix generation module, and a dynamic weighted multi-scale graph convolutional network.

[0019] As a further improvement of the present invention, the multi-head attention mechanism projects the time-frequency hybrid features into the multi-head attention space, and splices the output features generated by each attention head to obtain a spliced tensor.

[0020] As a further improvement of the present invention, in the adaptive adjacency matrix generation module, a graph attention network is introduced to capture the multi-scale dependence relationship between the spliced tensors and generate an adaptive adjacency matrix.

[0021] As a further improvement of the present invention, in the dynamic weighted multi-scale graph convolutional network, a learnable dynamic weight factor is introduced into the adaptive adjacency matrices of different orders, and the contribution degree of each graph convolution scale in feature fusion is adjusted by the learnable dynamic weight factor to generate multi-scale features.

[0022] As a further improvement of the present invention, project the multi-scale features back to the original feature space to generate a multi-scale tensor.

[0023] As a further improvement of the present invention, in the multi-scale multi-frequency fusion module, fuse the multi-scale tensor in the frequency scale and the time scale to obtain a fused feature, and the calculation formula is:

[0024]

[0025] where, F fusion represents the fused feature, f represents the ReLU non-linear activation function, k represents the index, K represents the set of all feature sources participating in feature fusion, K = {(m, l)|1 ≤ m ≤ M, 1 ≤ l ≤ L}, m represents the index of the time scale, l represents the index of the frequency component, X(k) Denote the feature with index k, w k Denote the weight assigned to each feature X (k) , where X represents the input feature vector.

[0026] As a further improvement of the present invention, the fused features are mapped to the battery life and health index space for battery state prediction.

[0027] A battery state prediction device based on a battery time series model, comprising:

[0028] A data processing module, which acquires a battery multivariate time series data set and preprocesses the battery multivariate time series data set;

[0029] A model construction module, which constructs a battery time series model, and the battery time series model includes a time-frequency feature extraction module, a multi-scale adaptive graph convolutional network, and a multi-scale multi-frequency fusion module;

[0030] A model processing module, which uses the time-frequency feature extraction module to extract the time-frequency hybrid features in the preprocessed battery multivariate time series data set; the multi-scale adaptive graph convolutional network captures the multi-scale features in the time-frequency hybrid features and processes the multi-scale features to obtain a multi-scale tensor; the multi-scale multi-frequency fusion module fuses the multi-scale tensor in the frequency scale and the time scale to obtain fused features;

[0031] A state prediction module, which predicts the battery state through the fused features.

[0032] 3. Beneficial effects

[0033] Compared with the prior art, the advantages of the present invention are as follows:

[0034] The battery state prediction method and device based on a battery time series model of the present invention combine the discrete Fourier transform with orthogonal basis functions to realize the extraction of time-frequency features of battery data, introduce an adaptive graph convolutional network in the multi-scale dimension, dynamically weight and fuse the complex dependence relationships between battery data. Thus, the decay features of the battery in multiple frequencies and multiple time scales can be accurately captured, and the accuracy and robustness of the remaining life prediction and health state assessment are improved. Description of the drawings

[0035] Figure 1 It is a schematic structural diagram of the battery time series model according to an embodiment of the present invention;

[0036] Figure 2 It is a working flowchart of the multi-scale adaptive graph convolutional network according to an embodiment of the present invention. Detailed implementation manners

[0037] The present invention will be described in detail below in conjunction with the drawings in the specification and specific embodiments.

[0038] Embodiment

[0039] A battery state prediction method based on a battery time series model provided in this embodiment includes the following steps: obtaining a battery multivariate time series dataset and preprocessing the battery multivariate time series dataset; constructing a battery time series model, which includes a time-frequency feature extraction module, a multi-scale adaptive graph convolutional network, and a multi-scale multi-frequency fusion module; using the time-frequency feature extraction module to extract the time-frequency mixed features in the preprocessed battery multivariate time series dataset; the multi-scale adaptive graph convolutional network captures the multi-scale features in the time-frequency mixed features and processes the multi-scale features to obtain a multi-scale tensor; the multi-scale multi-frequency fusion module fuses the multi-scale tensor in the frequency scale and the time scale to obtain a fusion feature; and predicting the battery state through the fusion feature.

[0040] Specifically in this embodiment, a battery multivariate time series dataset is obtained. The battery multivariate time series dataset includes parameters such as timestamps, battery voltages, currents, temperatures, and remaining useful life (RUL) of the battery. Further, the battery multivariate time series dataset is preprocessed. Specifically, the battery multivariate time series dataset is cleaned, formatted, and standardized to improve data quality. In addition, existing interpolation methods and clustering methods are used to fill in the missing values in the battery multivariate time series dataset, outliers are removed, and time alignment and resampling are performed on different sensor data. Existing feature engineering is used to generate statistical features, physical features, and time series features, and the performance of the battery time series model is improved through normalization and standardization. In the embedding process, a sliding window is used to extract time series subsequences, and position embeddings are combined to capture time series features to generate a unified and efficient multi-variable input representation, ensuring that the subsequent battery time series model can fully learn the dynamic behavior of the battery.

[0041] Further, as Figure 1 shown, a battery time series model is constructed. In this embodiment, the battery time series model includes a time-frequency feature extraction module, a multi-scale adaptive graph convolutional network, and a multi-scale multi-frequency fusion module. It should be noted that the battery time series model constructed in this embodiment can efficiently extract the potential laws in the multi-time scale and multi-frequency information of the power battery, and finally realize the accurate modeling of the remaining useful life (RUL) prediction, health state assessment, and capacity decay trend of the battery.

[0042] In this embodiment, a time-frequency feature extraction module is used to extract the time-frequency hybrid features from the preprocessed battery multi-temporal dataset. The specific steps are as follows: Perform a discrete Fourier transform on the preprocessed battery multi-temporal dataset to obtain a complex spectrum, which includes a real component and an imaginary component. The real component and the imaginary component respectively correspond to the amplitude and phase information of the frequency-domain signal at each frequency component. It should be noted that in this embodiment, the frequency-domain signal refers to the battery multi-temporal data used for analysis in the process of time-frequency hybrid feature extraction. After the battery multi-temporal data is Fourier-transformed, it is converted into a frequency-domain signal, and the real component and the imaginary component therein respectively contain the amplitude and phase information of the frequency-domain signal. In this embodiment, the preprocessed battery multi-temporal dataset is set as where N represents the number of channels and T represents the number of time steps. Then, for each channel of the preprocessed battery multi-temporal dataset X (norm) , perform a discrete Fourier transform to obtain a complex spectrum z ∈ C 1×T . The real component and the imaginary component of this complex spectrum respectively represent z R and z I . The calculation formula is:

[0043]

[0044] Z R = Re(z), z I = Im(z)

[0045] where z [k] represents the k-th component of the input sequence in the frequency-domain representation, k represents the frequency index, ranging from 0 to T - 1, corresponding to different frequency components, n represents the summation index, and its range is from 0 to T - 1, indicating the accumulation of each sampling point in the time domain. T represents the time length of the frequency-domain signal, and X (norm) [n] represents the preprocessed battery multi-temporal data, represents the complex exponential function, z R represents the real component, and z I represents the imaginary component.

[0046] In this embodiment, the real component z R corresponds to the amplitude information of different frequency components, and the imaginary component corresponds to the phase information of different frequency components. Thus, through the discrete Fourier transform, the preprocessed battery multi-temporal data can be converted from the time domain to the frequency domain, and the periodic components in the sequence can be explicitly separated.

[0047] Furthermore, the real component is multiplied by the orthogonal cosine basis function to obtain the time series feature, the imaginary component is multiplied by the orthogonal sine basis function to obtain the frequency domain feature, and the time series feature and the frequency domain feature are added to obtain the time-frequency hybrid feature. Specifically, in order to retain the phase information in the time series dimension and extract the time series information and the frequency domain information simultaneously, in this embodiment, the orthogonal cosine basis function and the orthogonal sine basis function with a time length of T of the frequency domain signal are introduced, where each frequency corresponds to a group of basis functions, which is expressed as:

[0048]

[0049] where C (k,n) represents the orthogonal cosine basis function, S(k,n) represents the orthogonal sine basis function, T represents the time length of the frequency domain signal, k represents the frequency index, k = 0, 1, 2, …, T / 2, and n represents the time index, n = 0, 1, 2, …, T-1.

[0050] Thus, in this embodiment, the real component is multiplied by the orthogonal cosine basis function and the imaginary component is multiplied by the orthogonal sine basis function and then added to obtain the time-frequency hybrid feature, and its calculation formula is:

[0051]

[0052] where G [n] represents the time-frequency hybrid feature, k represents the frequency index, T represents the number of time steps of the signal, z R [k] represents the real component of the complex spectrum, and z I [k] represents the imaginary component of the complex spectrum.

[0053] The amplitude information and phase information of the time series data at different frequencies will be incorporated into the time series dimension in an explicit manner, which not only retains the original time resolution but also avoids the ambiguity caused by inconsistent phases or mismatched sequence lengths. Thus, in this embodiment, the real component and the imaginary component of the complex spectrum are respectively expanded by the basis functions using the orthogonal cosine basis function and the orthogonal sine basis function, and the complex spectrum obtained by the discrete Fourier transform is mapped onto the orthogonal cosine and sine bases, so as to explicitly retain and express the phase information of the frequency domain signal in the time domain, while taking into account the capture of periodicity and global trends in the frequency domain. Finally, the two are added to obtain the time-frequency hybrid feature, and the time-frequency hybrid feature carries the phase information and amplitude information of different frequency components in the time dimension, and can better reveal the periodic changes and potential laws in the multivariate time series.

[0054] Furthermore, the multi-scale adaptive graph convolutional network captures the multi-scale features in the time-frequency hybrid feature and processes the multi-scale features to obtain the multi-scale tensor. Specifically, as Figure 2As shown in the figure, the multi-scale adaptive graph convolutional network includes a multi-head attention mechanism, an adaptive adjacency matrix generation module, and a dynamic weighted multi-scale graph convolutional network.

[0055] It should be noted that the discrete Fourier transform and the orthogonal basis function expansion obtain the time-frequency mixed features, and the time-frequency mixed features comprehensively retain the information of the frequency-domain signal in the time domain and the frequency domain in the form of a tensor. To enhance the ability to capture cross-time series dependencies, in this embodiment, a multi-head attention mechanism is introduced, and the time-frequency mixed feature tensor is used as the input tensor. Then, the input tensor is projected into the multi-head attention space, and then the output tensors generated by each attention head are concatenated to form a concatenated tensor. Its calculation formula is expressed as:

[0056] Y i = Concat(head 1 , head 2 ,..., head k )W o

[0057]

[0058] Among them, Y i represents the concatenated tensor, Concat represents the concatenation operation, head k represents the output of the k-th attention head, k represents the number of attention heads, W o represents the output linear transformation matrix, Softmax represents the Softmax activation function, Q k represents the query matrix, K k represents the key matrix, V k represents the value matrix, T represents the transpose operation, and d k represents the dimension size of each query, key, and value.

[0059] Furthermore, in this embodiment, in the adaptive adjacency matrix generation module, a Graph Attention Network (GAT) is introduced to capture the multi-scale dependencies between the concatenated tensors and generate an adaptive adjacency matrix. It should be noted that in this embodiment, the graph attention network is a graph neural network structure based on the attention mechanism, which can update and generate the adjacency matrix by learning the attention weights between nodes, so as to perform more flexible feature aggregation on the graph structure. Thus, through the graph attention network, the relationships between nodes in the preprocessed battery multi-temporal dataset are dynamically modeled, so that the adjacency matrix can adaptively reflect the correlation degree between different nodes, and further improve the ability of subsequent multi-scale graph convolution to capture complex dependencies. In this embodiment, the multi-scale dependencies between the output tensors are captured through the graph attention network, and an adaptive adjacency matrix is generated. Its calculation formula is:

[0060]

[0061] Among them, represents the adaptive adjacency matrix, which is used to reflect the correlation strength between time series. i represents the source node in the graph, j represents the target node in the graph, LeakyReLU represents the LeakyReLU activation function, α represents the learnable attention vector, and α T represents the transpose of the attention weight vector. and represent different trainable node embedding matrices, which are used to represent the low-dimensional feature representations of nodes, and || represents the feature concatenation operation.

[0062] It should be noted that in this embodiment, in order to effectively fuse the multi-scale dependence relationships in the preprocessed battery multivariate time series dataset, a dynamic weighted multi-scale graph convolution method is proposed, including: by introducing learnable dynamic weight factors into the adaptive adjacency matrices of different orders, the contribution degrees of each graph convolution scale in feature fusion can be adaptively adjusted, so as to take into account both short-range local information and long-range global dependence information. In this process, the battery time series model can not only be more robust under noise interference and abnormal data distribution conditions, but also flexibly capture complex features such as periodicity, trend, and mutation of the battery under various dynamic working conditions, thus significantly improving the accuracy and stability of battery remaining life prediction and health state assessment. Through the dynamic weighted multi-scale graph convolution network, the information loss and insufficient dependence caused by a single scale are effectively overcome, and the ability of the battery time series model to mine multi-scale features of battery multivariate time series data is enhanced. In this embodiment, in the dynamic weighted multi-scale graph convolution network, learnable dynamic weight factors are introduced into the adaptive adjacency matrices of different orders, and the contribution degrees of each graph convolution scale in feature fusion are adjusted by the learnable dynamic weight factors to generate multi-scale features. The multi-scale features are represented by the adaptive adjacency matrices of different orders, and their fusion process is expressed as:

[0063]

[0064] Among them, represents the output tensor, σ represents the activation function, j represents the index of the matrix node, p represents the set of orders of the adaptive adjacency matrix, and α j represents the dynamic weight, which is calculated by the attention mechanism, and its calculation formula is:

[0065]

[0066] e j = w T ReLU((A i ) j Y i )

[0067] Among them, w represents learnable parameters, which are used to dynamically model the importance of multi-scale features.

[0068] After the multi-scale graph convolution is completed, the multi-scale features obtained by each layer of the network are high-dimensional embedding representations with different scale information. Therefore, it is necessary to project the multi-scale features back to the original feature space to generate a multi-scale tensor. In this embodiment, the original feature space is the feature representation space of the battery multivariate time series dataset before being processed by the deep learning model. These original features include multiple time series data parameters such as the voltage, current, temperature, remaining useful life (RUL), etc. of the battery. Projecting the multi-scale features back to the original feature space so that the dimension is consistent with the original battery multivariate time series data helps to ensure that the output features are consistent with the requirements of the battery state prediction task at the output stage of the battery time series model. Especially when dealing with practical applications such as remaining useful life prediction and health state assessment, it can avoid the misalignment problem caused by the inconsistency between the high-dimensional features of the battery time series model and the task target dimension.

[0069] Specifically, projecting the multi-scale features back to the original feature space makes the dimension consistent with the target dimension of the subsequent prediction task at the final output stage. Through this projection process, not only can the key features extracted by the multi-scale convolution be retained, but also the output of the battery time series model can be matched with the original input or the requirements of the target task, avoiding the misalignment problem between the high-dimensional embedding and the actual application scenario. In this embodiment, the formula for projecting the multi-scale features back to the original feature space is:

[0070]

[0071] Among them, represents the i-th feature projected back to the original feature space, and MLP represents a multi-layer perceptron, which includes one or more fully connected layers and combines a non-linear activation function to enhance the expression ability of the output features.

[0072] Therefore, in this embodiment, the multi-scale adaptive graph convolution network combines the multi-head attention mechanism, adaptive adjacency matrix generation, and dynamic weighted multi-scale graph convolution, and can effectively capture the spatio-temporal features of complex time series data, providing higher prediction accuracy and robustness.

[0073] After the multi-scale adaptive graph convolution network is completed, a multi-scale tensor is obtained. This multi-scale tensor aggregates the embedding representations of the relationships between different time scales and multi-dimensional features in multiple layers of the network. To make this multi-scale tensor better highlight the key features and suppress the useless noise in the subsequent prediction tasks, in this embodiment, the multi-scale tensor is fused on the frequency scale and the time scale to obtain a fused feature, and the formula is:

[0074]

[0075] Among them, F fusion represents the fused feature, f represents the ReLU non-linear activation function, k represents the index used to traverse different feature sources participating in the fusion, K represents the set of all feature sources participating in feature fusion, K = {(m, l)|1 ≤ m ≤ M, 1 ≤ l ≤ L}, m represents the index of the time scale, l represents the index of the frequency component, X (k) represents the feature with index k, w k represents the weight assigned to each feature X (k) , and X represents the input feature vector.

[0076] Therefore, the fused feature is mapped to the battery life and health index space for battery state prediction. In this embodiment, the remaining useful life (RUL) prediction and the result of the capacity attenuation trend can be directly generated at the output layer of the battery time series model. On the one hand, the time-frequency hybrid features and the deep features extracted by the multi-scale graph convolutional network are aligned with the known life, capacity, or health state labels and supervised training is carried out, so as to realize the accurate mapping from complex working condition multivariate data to attenuation and life indicators; on the other hand, the introduction of dynamic weighting and the adaptive adjacency matrix effectively enhances the adaptability of the battery time series model to the battery operating environment and multi-dimensional correlations, enabling accurate and stable remaining useful life, capacity degradation trend, and health state assessment to be quickly given during actual inference, providing more targeted decision-making basis for the monitoring and early warning of the battery management system.

[0077] A battery state prediction method based on a battery time series model provided by this embodiment combines the traditional discrete Fourier transform with orthogonal basis functions, so as to explicitly retain the phase information in the time domain and generate a hybrid representation with both time and frequency characteristics. The time domain representation retains the time order and local dynamic information of the signal, while the frequency domain representation is more adept at capturing periodicity and global trends. Through this hybrid representation method, the information of the signal in the time and frequency dimensions can be comprehensively reflected in a unified feature space, thus more comprehensively revealing the multi-scale dynamic changes and periodic characteristics of battery data. In addition, the multi-scale adaptive graph convolutional network can effectively capture the multi-scale dependence relationships and spatio-temporal characteristics in the power battery time series data, providing higher prediction accuracy and robustness.

[0078] This embodiment also provides a battery state prediction device based on a battery time series model, including a data processing module, a model construction module, a model processing module, and a state prediction module. The data processing module is used to obtain and input a battery multivariate time series data set and preprocess the battery multivariate time series data set. The model construction module is used to construct a battery time series model, which includes a time-frequency feature extraction module, a multi-scale adaptive graph convolutional network, and a multi-scale multi-frequency fusion module. The model processing module uses the time-frequency feature extraction module to extract the time-frequency mixed features in the preprocessed battery multivariate time series data set; the multi-scale adaptive graph convolutional network captures the multi-scale features in the time-frequency mixed features and processes the multi-scale features to obtain a multi-scale tensor; the multi-scale multi-frequency fusion module fuses the multi-scale tensor in the frequency scale and the time scale to obtain a fused feature. The state prediction module predicts the battery state through the obtained fused feature. The battery state prediction device based on a battery time series model provided in this embodiment can implement any of the methods of the battery state prediction method based on a battery time series model, and the specific working process of the battery state prediction device based on a battery time series model can refer to the corresponding process in the embodiment of the battery state prediction method based on a battery time series model. The methods and devices provided in this embodiment can be implemented in other ways. For example, the device embodiments described above are merely illustrative; for example, the division of a certain module is only a logical function division, and there may be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the connections or communication connections shown or discussed with each other can be indirect couplings or communication connections through some interfaces, devices, or units, and can also be electrical, mechanical, or other forms of connections.

[0079] This embodiment also provides a computer device. A computer device includes a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the computer program, it implements the battery state prediction method based on a battery time series model described above.

[0080] This embodiment also provides a computer-readable storage medium. A computer-readable storage medium stores a computer program, and when the computer program is run by a processor, it executes the battery state prediction method based on a battery time series model described in this embodiment. Among them, the computer-readable storage medium can be any tangible medium that contains or stores a program, and this program can be used by or in combination with an instruction execution system, device, or device; the program code contained on the computer-readable medium can be transmitted by any appropriate medium, including but not limited to wireless, wire, optical fiber, RF, etc., or any suitable combination of the above.

[0081] The present invention and its implementation manners are schematically described above. The description is not restrictive. Without departing from the spirit or basic features of the present invention, the present invention can be implemented in other specific forms. What is shown in the drawings is only one of the implementation manners of the present invention, and the actual structure is not limited thereto. Any reference signs in the claims should not limit the claimed claims. Therefore, if those of ordinary skill in the art are inspired by it and, without departing from the gist of the present invention, design similar structural manners and embodiments to the technical solution without creative efforts, they shall fall within the protection scope of the present invention. In addition, the term "comprising" does not exclude other elements or steps, and the word "a" before an element does not exclude including "a plurality of" such elements. The plurality of elements stated in the product claims can also be implemented by one element through software or hardware. The terms such as "first" and "second" are used to indicate names and do not indicate any specific order.

Claims

1. A battery state prediction method based on a battery timing model, comprising the following steps: Acquire a battery multivariate time series data set, and preprocess the battery multivariate time series data set; Constructing a battery timing model, wherein the battery timing model includes a time-frequency feature extraction module, a multi-scale adaptive graph convolutional network, and a multi-scale multi-frequency fusion module; The time-frequency feature extraction module is used to extract the time-frequency mixed features in the preprocessed battery multivariate time series data set; The multi-scale adaptive graph convolutional network captures the multi-scale features in the time-frequency mixed features and processes the multi-scale features to obtain multi-scale tensors; the multi-scale multi-frequency fusion module fuses the multi-scale tensors on the frequency scale and time scale to obtain fused features; The battery status is predicted by fusing features.

2. A battery state prediction method based on a battery timing model according to claim 1, characterized in that: The method of using the time-frequency feature extraction module to extract the time-frequency mixed features in the preprocessed battery multivariate time series data set specifically includes the following steps: Performing a discrete Fourier transform on the preprocessed battery multivariate time series data set to obtain a complex spectrum, wherein the complex spectrum includes a real component and an imaginary component; The real component is multiplied by the orthogonal cosine basis function to obtain the time series feature, the imaginary component is multiplied by the orthogonal sine basis function to obtain the frequency domain feature, and the time series feature and the frequency domain feature are added to obtain the time-frequency mixed feature.

3. A battery state prediction method based on a battery timing model according to claim 2, characterized in that: The multi-scale adaptive graph convolutional network includes a multi-head attention mechanism, an adaptive adjacency matrix generation module and a dynamic weighted multi-scale graph convolutional network.

4. The battery state prediction method based on the battery timing model according to claim 3 is characterized in that: The multi-head attention mechanism projects the time-frequency mixed features into the multi-head attention space, and concatenates the output features generated by each attention head to obtain a concatenated tensor.

5. The battery state prediction method based on the battery timing model according to claim 4 is characterized in that: In the adaptive adjacency matrix generation module, a graph attention network is introduced to capture the multi-scale dependencies between concatenated tensors and generate an adaptive adjacency matrix.

6. A battery state prediction method based on a battery timing model according to claim 5, characterized in that: In the dynamic weighted multi-scale graph convolutional network, learnable dynamic weight factors are introduced into adaptive adjacency matrices of different orders. The contribution of each graph convolution scale in feature fusion is adjusted by the learnable dynamic weight factors to generate multi-scale features.

7. A battery state prediction method based on a battery timing model according to claim 6, characterized in that: The multi-scale features are projected back to the original feature space to generate a multi-scale tensor.

8. A battery state prediction method based on a battery timing model according to claim 7, characterized in that: In the multi-scale multi-frequency fusion module, multi-scale tensors are fused on the frequency scale and time scale to obtain fusion features. The calculation formula is: Among them, F fusion represents the fusion feature, f represents the ReLU nonlinear activation function, k represents the index, K represents the set of all feature sources involved in feature fusion, K={(m,l)|1≤m≤M,1≤l≤L}, m represents the index of the time scale, l represents the index of the frequency component, X (k) represents the feature with index k, w k Represents the assignment to each feature X (k) The weight of X represents the input feature vector.

9. A battery state prediction method based on a battery timing model according to claim 8, characterized in that: The fused features are mapped to the battery life and health indicator space for battery status prediction.

10. A battery state prediction device based on a battery timing model, characterized in that: include: A data processing module, which obtains a battery multivariate time series data set and preprocesses the battery multivariate time series data set; A model building module, which builds a battery timing model, wherein the battery timing model includes a time-frequency feature extraction module, a multi-scale adaptive graph convolutional network, and a multi-scale multi-frequency fusion module; The model processing module uses the time-frequency feature extraction module to extract the time-frequency mixed features in the preprocessed battery multivariate time series data set; The multi-scale adaptive graph convolutional network captures the multi-scale features in the time-frequency mixed features and processes the multi-scale features to obtain multi-scale tensors; the multi-scale multi-frequency fusion module fuses the multi-scale tensors on the frequency scale and time scale to obtain fused features; The state prediction module predicts the battery state by fusing features.

Images