Lithium battery SOC adaptive prediction method based on multi-scale map convolutional network

Abstract

The invention provides a lithium battery SOC adaptive prediction method based on a multi-scale map convolutional network, and the method comprises the steps: 1, obtaining monitoring data of a lithium battery, carrying out the data cleaning and standardization processing of the monitoring data, and obtaining an input feature sequence Xinput at a current moment; step 2, inputting the current time input feature sequence Xinput into a lithium battery SOC adaptive prediction network based on a multi-scale map convolutional network, and outputting a lithium battery SOC prediction result; wherein the built lithium battery SOC adaptive prediction network based on the multi-scale image convolution network comprises a multi-scale feature extractor, a spatial relation feature extractor and a prediction module. According to the method, the spatial dependency relationship among the parameters is captured by using the graph convolutional network, the characteristics under different time scales are extracted by combining a multi-scale analysis technology in order to capture the multi-scale dynamic change in the charging and discharging process of the battery, so that the accuracy and robustness of SOC prediction are improved, the performance of a lithium battery management system is improved, and the service life of the battery is prolonged.

Description

Technical Field

[0001] The present invention belongs to the field of artificial intelligence technology, and specifically relates to a lithium battery SOC adaptive prediction method based on a multi-scale graph convolutional network. Background Art

[0002] Lithium batteries play an increasingly important role in modern society, particularly in electric vehicles, energy storage systems, and portable electronic devices. As reliance on lithium batteries increases, so too do the requirements for their performance and safety. State of charge (SOC) prediction (SOC) is a key technology for ensuring safe and efficient operation of lithium batteries. The accuracy of SOC prediction directly impacts the performance of the battery management system, and thus the efficiency and lifespan of the entire battery system.

[0003] Traditional SOC prediction methods are primarily based on mathematical models, such as the open-circuit voltage method, coulomb counter method, and Kalman filter method. While these methods can achieve a certain degree of SOC prediction, they have limitations in practical applications due to the nonlinear and time-varying characteristics of lithium batteries. For example, the open-circuit voltage method requires the battery to be in a static state for accurate measurement. However, in practical applications, batteries are often in a dynamic operating state, leading to limitations in the open-circuit voltage method. Similarly, while the coulomb counter method can monitor the battery's charge and discharge status in real time, its accuracy is also limited by factors such as battery self-discharge and temperature fluctuations. The Kalman filter method, as a dynamic estimation method, can overcome these issues to a certain extent, but its complex calculation process and high requirements for model accuracy also limit its widespread use in practical applications.

[0004] With the development of artificial intelligence technology, SOC prediction methods based on deep learning have gradually become a research hotspot. Deep learning methods can learn the charging and discharging characteristics of batteries through large amounts of data, thereby achieving accurate prediction of SOC. For example, methods based on recurrent neural networks (RNNs) can effectively capture the time series characteristics of the battery charging and discharging process, methods based on convolutional neural networks (CNNs) can extract the spatial characteristics of parameters such as battery voltage and current, and methods based on long short-term memory networks (LSTMs) can better handle the long-term dependencies in the battery charging and discharging process. However, most deep learning methods only consider the characteristics of the time series and ignore the spatial dependencies between battery parameters, which in turn limits the performance of the prediction model. For example, the battery charging and discharging process is related to parameters such as battery temperature and internal resistance, and there are complex spatial dependencies between these parameters. If these spatial dependencies are not considered, it will be difficult to achieve accurate prediction of SOC. Summary of the Invention

[0005] The purpose of the present invention is to solve the problem that the existing SOC prediction method ignores the spatial dependency between battery parameters, and provides a lithium battery SOC adaptive prediction method based on a multi-scale graph convolutional network. By constructing a graph representation of battery parameters and using a graph convolutional network to capture the spatial dependency between parameters, in order to capture the multi-scale dynamic changes in the battery charging and discharging process, the multi-scale characteristics of the lithium battery are comprehensively considered. By combining multi-scale analysis technology, features at different time scales are extracted to better capture the dynamic changes in the battery charging and discharging process, effectively improving the accuracy and robustness of SOC prediction, improving the performance of the lithium battery management system and extending the battery life.

[0006] To achieve the above objectives, the technical solutions provided by the present invention are:

[0007] A lithium battery SOC adaptive prediction method based on a multi-scale graph convolutional network, comprising:

[0008] Step 1: Obtain the monitoring data of the lithium battery and perform data cleaning and standardization to obtain the current input feature sequence X input ;

[0009] Step 2: Input the current moment into the feature sequence X input Input to the lithium battery SOC adaptive prediction network based on the multi-scale graph convolutional network, and output the lithium battery SOC prediction result; wherein: the constructed lithium battery SOC adaptive prediction network based on the multi-scale graph convolutional network includes: a multi-scale feature extractor, a spatial relationship feature extractor and a prediction module, specifically:

[0010] The multi-scale feature extractor is designed based on the separation of large convolution kernels and multi-path parallel topology architecture. The multi-scale feature extractor extracts the current moment input feature sequence X obtained in step 1 through the multi-scale feature extraction module. input Processing is performed and multi-path depth separation convolution operation is used to obtain output features;

[0011] The features extracted by the graph convolutional network in the spatial relationship feature extractor contain rich spatial information, reflecting the current state of each node and the information fusion of adjacent nodes, thereby improving the representation ability of node features;

[0012] The prediction module predicts the SOC of the lithium battery to obtain a lithium battery SOC prediction result.

[0013] As a further improvement of the present invention, in step 2, (11) the multi-scale feature extractor is designed based on a separate large convolution kernel and a multi-path parallel topology architecture, including: utilizing a channel decoupling mechanism to perform a convolution transformation on the selected channels and retaining the remaining channels as identity mapping branches; specifically, it includes:

[0014] (111) By feeding the inputs of different branches into different parallel branches, the convolution operation of each branch can be expressed as:

[0015]

[0016] In formula (1), X′ hw Represents branch X hw Select the depth-separable convolution with convolution kernel size of 3×3 and 1×1 to extract effective features of the input data, X′ W Represents branch X w Select the depth-separable convolution with convolution kernel size of 1×7 and 1×1 to extract effective features of the input data, X′ h Represents branch X h Select the depth-separable convolution with convolution kernel size of 7×1 and 1×1 to extract effective features of the input data, X′ id Represents branch X id A depth-wise separable convolution with a convolution kernel size of 1×1 is used to extract effective features from the input data. The 1×1 convolution kernel is used to integrate channel information for feature combination and enhanced nonlinear mapping, the 3×3 convolution kernel is used for local spatial correlation, the 1×7 convolution kernel is used to extract long-range dependencies in the time dimension, and the 7×1 convolution kernel is used to extract cross-channel features.

[0017] (112) Based on the separated large convolution kernel, the outputs of each branch are connected to obtain the connected output X′, which is expressed as:

[0018] X′=Concat(X′ hw ,X′ w ,X′ h ,X′ id ) Formula (2)

[0019] Regularization operation is performed on the connected output X′, and the expression is:

[0020] Y=Norm(X′) Formula (3)

[0021] In formula (3), Y represents the operation result;

[0022] The multi-scale feature extractor uses residual connections to improve gradient fluidity and network convergence speed during training, which can be expressed as:

[0023]

[0024] In formula (4), Y represents the operation result, r represents the expansion coefficient, C represents the input channel size, σ represents the activation function, and X represents the data set;

[0025] (12) The multi-scale feature extraction module in step 2 extracts the current moment input feature sequence X obtained in step 1. input Processing, including:

[0026] The multi-scale feature extractor is used to extract the current input feature sequence X obtained in step 1. input , group them according to the channel dimension, and divide the channel dimension into two subsets, which are expressed as:

[0027] X hw ,X w ,X h ,X id =Split(X)

[0028] =X :,:g ,X :g:2g ,X :2g:3g ,X :3g: Formula (5)

[0029] In formula (5), X hw represents the extracted features related to height and width, X w represents the grouping feature assigned to the horizontal band kernel, X h represents the grouping feature assigned to the vertical band kernel, X id Indicates the extraction of identity or channel related features, X :,:g represents the first g columns of the input tensor X, X :g:2g Represents the gth to 2gth columns of the input tensor X, X :2g:3g represents the grouping feature assigned to the vertical band kernel, X :3g: Represents the 3gth to the last column of the input tensor X;

[0030] The multi-scale feature extractor is constructed by setting a ratio r n Determine the number of each branch channel, expressed as n = r n C, where n represents the number of channels of the convolution branch and C represents the number of channels of the input tensor;

[0031] (13) In step 2, a multi-path depth separation convolution operation is used to obtain output features; wherein the output feature matrix X contained in the output feature is out1 Expressed as:

[0032] X out1 ∈R 4×P Formula (6)

[0033] In formula (6), R represents the expansion coefficient and P represents the dimension of the hidden layer;

[0034] The multi-scale feature extraction module is used to input the feature sequence X at the current momentinput The output features are captured through multi-channel decomposition.

[0035] As a further improvement of the present invention, the features extracted by the graph convolutional network in the spatial relationship feature extractor in step 2 contain rich spatial information, reflect the current state of each node and the information fusion of adjacent nodes, and enhance the representation capability of node features; including:

[0036] (21) Construct a feature interaction graph to describe the relationship between battery parameters, expressed as: G = (V, E), where G represents the feature interaction graph, V represents the node, and E represents the set of edges;

[0037] Each battery feature in the output features is regarded as a node V, expressed as: V = {I, V, T, SOC} ∈ R 4×n , where I represents the identity matrix, T represents time, SOC represents the lithium battery SOC prediction, and 4 represents the feature dimension of each time step;

[0038] The physical correlation between features is represented by the adjacency matrix A, which is expressed as: A∈R 4×4 , A i,j ∈{0,1}, where A i,j Indicates whether there is an explicit coupling relationship between feature i and feature j;

[0039] Construct feature matrix X 4×P , select the current moment input feature sequence X in step 1 input After window division, it is used as the feature matrix X 4×P , the multi-dimensional time series monitoring data of the battery system is used as the node feature X, expressed as: X∈R 4×P ;

[0040] The current, voltage, temperature, and capacity states are constructed as nodes in a graph structure. The physical coupling relationship between parameters is captured through graph convolution operations to obtain a graph convolutional network, which enables interactive modeling of multi-dimensional spatial features.

[0041] (22) The graph convolution operation corresponding to the graph convolution network is expressed as:

[0042]

[0043] In formula (7), X (l+1) Represents the node feature matrix of the l+1 layer, which represents the features of all nodes at the current level, X (l) represents the node feature matrix of the lth layer, σ represents the activation function, represents the degree matrix, Indicates that the adjacency matrix A is increased by the identity matrix I, W (l) represents the learnable weight matrix of layer l;

[0044] (23) After multi-layer graph convolution operations, the final output is the same as X out1 Feature matrices of the same dimension is the node feature enhanced by spatial features, expressed as: Where L represents the number of graph convolution layers and P represents the dimension of the hidden layer.

[0045] As a further improvement of the present invention, in step 2, the lithium battery SOC adaptive prediction network based on the multi-scale graph convolutional network constructed further includes at least: a time series feature extractor and a feature weighted fusion module, specifically:

[0046] The time series feature extractor uses a multi-scale multi-head self-attention mechanism to capture the sequence correlation within each time scale, and uses dimension redefinition to reshape the tensor of each scale into a two-dimensional matrix;

[0047] The feature fusion weighting module integrates multi-scale features, spatial relationship features and time series features into a unified framework, and fully integrates them through the weight parameters of the scientific system to obtain fusion features;

[0048] Preferably, during the training process of the lithium battery SOC adaptive prediction network based on the multi-scale graph convolutional network, the fusion features are input into the prediction module to predict the SOC of the lithium battery, and the various parameters of the lithium battery SOC adaptive prediction network are adjusted to obtain the lithium battery SOC prediction result.

[0049] As a further improvement of the present invention, the time series feature extractor uses fast Fourier transform to detect the current input feature sequence X input The periodic characteristics and assigning an appropriate time scale to each period are expressed as:

[0050]

[0051] In formula (8), FFT(·) represents the input feature sequence X at the current moment input For FFT calculation, Amp(·) represents the calculation of amplitude, F represents a vector, and the amplitude of each frequency reflects the periodic characteristics on different time scales, f1,…,f k Represents the frequency amplitude vector, whose dimension is the same as the number of frequency components after Fourier transform, and each element represents the amplitude value of the corresponding frequency, s i represents the frequency index, f i represents the i-th frequency amplitude vector, and L represents the number of graph convolution layers;

[0052] Based on the selected time scale {s1,…,s k}, the input data is reshaped into a 3D tensor by the following formula to obtain several representations corresponding to different time scales, the expression is:

[0053]

[0054] In formula (9), Padding(·) is used to extend the time series by zero along the time dimension, X (l) represents the reshaped feature matrix, represents the i-th reshaped time series based on time scale i, Represents the dimension reshaping function, the first dimension of the reshaped tensor is s i , the second dimension is f i ;

[0055] For the feature tensor F after Fourier transform processing, the multi-head self-attention mechanism is applied on its time scale dimension to extract the time correlation features, which can be expressed as:

[0056]

[0057] In formula (10), Represents the tensor of each scale, MHA S (·) represents the multi-head attention function in the scale dimension;

[0058] The tensor of each scale Shaped into a two-dimensional matrix The tensors are weighted and aggregated according to the amplitude of each scale, and the frequency values ​​in the Fourier transform results are Apply the SoftMax function to calculate the amplitude, the expression is:

[0059]

[0060] In formula (11), represents the output of the SoftMax function, represents the normalized weights of k branches, represents the final weighted fusion output, represents the i-th normalized weight, represents the independent output of k branches;

[0061] Features at different scales are weighted and aggregated according to their amplitudes, emphasizing significant information at each scale. By calculating the amplitude, the time series feature extractor is helped to adaptively focus on the most relevant temporal features based on the periodicity strength of each scale. Based on the expert mixture strategy, the time series feature extractor is enabled to efficiently integrate multi-scale time series data and extract more comprehensive and detailed time-dependent features.

[0062] Output time correlation feature Xout3 , denoted as X out3 ∈R Px4 , in order to enhance the performance of the time series feature extractor in multi-scale time series modeling, capture the temporal correlation at different scales, and better cope with lithium battery SOC prediction.

[0063] As a further improvement of the present invention, the feature fusion module fuses the multi-scale features extracted by the multi-scale feature extractor, the spatial relationship features extracted by the spatial relationship feature extractor, and the time series features extracted by the time series feature extractor in the information fusion stage, and outputs the fused features using a fully connected layer; specifically:

[0064] The dynamic weight learning mechanism further enhances the flexibility of fusion. It introduces learnable attention weights to adapt to the characteristic changes of different battery charging and discharging stages, and automatically adjusts the fusion ratio of multi-scale, spatial, and temporal features according to the distribution of input data.

[0065] The multi-scale features, spatial relationship features and time series features are fused using the weighted summation method. Assume that the feature vectors of the three scales are the multi-scale features X out1 , spatial relationship feature X out2 and time series features X out3 , whose corresponding weight coefficients are w1, w2, w3 respectively, and the feature vector X in the fusion feature after fusion by the feature weighted fusion module out Expressed as:

[0066] X out =w1·X out1 +w2·X out2 +w3·X out3 Formula (12)

[0067] The first weight coefficient w1, the second weight coefficient w2, and the third weight coefficient w3 are obtained through learning, and the contribution ratio of each scale feature is automatically adjusted according to the training process to achieve more flexible feature fusion.

[0068] As a further improvement of the present invention, in the prediction stage, step 2 inputs the fusion features into a fully connected layer in the prediction module, and the fully connected layer adjusts its output through an activation function to generate the final prediction result; the expression of the fully connected layer is:

[0069] Y=Sigmoid(Linear(W·X out +b)) Formula (12)

[0070] In formula (12), Sigmoid represents the activation function, which is used to increase nonlinearity and improve the expressiveness of the model, W represents the weight matrix of the fully connected layer, Xout represents the fused feature vector, and b represents the bias vector of the fully connected layer.

[0071] As a further improvement of the present invention, the lithium battery SOC adaptive prediction network based on the multi-scale graph convolutional network is trained, including:

[0072] Construct and preprocess dataset D t , and for this data set D t Divide into training set, validation set and test set; among them, the data set D t Expressed as:

[0073] The lithium battery SOC adaptive prediction network is trained and optimized using the training set and test set. The optimizer is used to update and adjust the model parameters. During the training process, the performance on the validation set is continuously monitored and the parameters are adjusted to prevent overfitting. The trained SOC adaptive prediction network is obtained.

[0074] After training, the test set is input into the trained lithium battery SOC adaptive prediction network for testing, and its prediction performance and stability are evaluated to accurately predict SOC.

[0075] The advantages of the present invention are:

[0076] 1. The present invention effectively captures battery features at different time scales through a multi-scale feature extraction module and provides a comprehensive feature representation. The present invention learns the spatial relationship between internal components of the battery based on a biased graph convolutional network, considers the bias term to handle missing values, and improves the robustness of the model.

[0077] 2. The present invention captures the time dependency of battery data through a multi-scale multi-head attention mechanism to improve the time series modeling effect. In the information fusion stage, the temporal dynamics, spatial structure and multi-scale characteristics of battery data are comprehensively considered, and the output layer is used for SOC prediction.

[0078] 3. The model training and evaluation stages of the present invention use mean square error as the loss function, optimize parameters through the gradient descent algorithm, and use the cross-validation method to evaluate the generalization ability of the model, so as to enable the lithium battery SOC prediction method to accurately predict the battery state, provide reliable decision support for the battery management system, and have good prediction accuracy and generalization ability.

[0079] Additional aspects and advantages of the present invention will be set forth in part in the description which follows and, in part, will be obvious from the description which follows, or may be learned by practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0080] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments with reference to the accompanying drawings, in which:

[0081] Figure 1 : Flowchart of a lithium battery SOC adaptive prediction method based on a multi-scale graph convolutional network provided by the present invention;

[0082] Figure 2 : A diagram of the network model of a lithium battery SOC adaptive prediction method based on a multi-scale graph convolutional network provided by the present invention when it is actually applied;

[0083] Figure 3 : Schematic diagram of the multi-scale feature extraction module provided by the present invention;

[0084] Figure 4 : Schematic diagram of the spatial relationship feature extraction module provided by the present invention;

[0085] Figure 5 : Schematic diagram of the temporal relationship feature extraction module provided by the present invention;

[0086] Figure 6 : The network structure and parameter setting table provided by the present invention;

[0087] Figure 7 : The simulation experiment result evaluation table provided by the present invention;

[0088] Figure 8 : Graphical representation of the original data curve and the predicted curve provided by the present invention. DETAILED DESCRIPTION

[0089] The following describes in detail embodiments of the present invention. The embodiments are exemplary and intended to explain the present invention, but are not to be construed as limiting the present invention.

[0090] See also Figure 1 , an embodiment of the present invention provides a lithium battery SOC adaptive prediction method based on a multi-scale graph convolutional network, comprising:

[0091] Step 1: Obtain the monitoring data of the lithium battery and perform data cleaning and standardization to obtain the current input feature sequence X input ;

[0092] Step 2: Input the current moment into the feature sequence X input Input to the lithium battery SOC adaptive prediction network based on the multi-scale graph convolutional network, and output the lithium battery SOC prediction result; wherein: the constructed lithium battery SOC adaptive prediction network based on the multi-scale graph convolutional network includes: a multi-scale feature extractor, a spatial relationship feature extractor and a prediction module, specifically:

[0093] The multi-scale feature extractor is designed based on the separation of large convolution kernels and multi-way parallel topology architecture. The multi-scale feature extractor uses the multi-scale feature extraction module to extract the current moment input feature sequence X obtained in step 1. input Processing is performed and multi-path depth separation convolution operation is used to obtain output features;

[0094] The features extracted by the graph convolutional network in the spatial relationship feature extractor contain rich spatial information, reflecting the current state of each node and the information fusion of adjacent nodes, improving the representation ability of node features;

[0095] The prediction module predicts the SOC of the lithium battery and obtains the lithium battery SOC prediction result.

[0096] More specifically, in the above step 1 of the embodiment of the present invention, the monitoring data of the lithium battery is obtained and the data is cleaned and standardized, including: collecting historical monitoring data of the lithium battery through experiments, including parameters such as voltage, current, temperature, and the battery's state of charge and discharge (SOC), and recording lithium battery data under different working conditions. These lithium battery data are organized into a time series format to form an initial data set. The constructed data set can be expressed as In addition, the battery capacity is determined as the label of the input data Therefore, the fully constructed dataset is The dataset is preprocessed, including data cleaning. The lithium battery data is cleaned to remove outliers and noise; normalization is performed to scale the data to a range suitable for model input; and finally, the dataset is divided into training set, validation set, and test set.

[0097] In practical applications, the aforementioned data preprocessing is the first step in lithium battery SOC prediction, aiming to ensure data quality and consistency, providing a reliable foundation for subsequent feature extraction and model training. Datasets typically contain multidimensional measurements of lithium batteries at different time points, such as voltage, current, temperature, and charge / discharge status. During the data preprocessing stage, the lithium battery data is first normalized and cleaned to remove outliers and noise, ensuring data quality. Furthermore, time window segmentation is required, dividing the continuous time series data into multiple windows, each containing a time series of a certain length for model input. Finally, the dataset is divided into training, validation, and test sets. The ratio can be adjusted based on actual conditions, with a typical split of 70% training, 15% validation, and 15% test.

[0098] Preferably, (11) the feature sequence X is input at the current moment inputThe acquisition process includes: obtaining battery monitoring characteristic data and battery capacity data by charging and discharging the battery in cycles. The test uses a charging and discharging cabinet as the main test equipment and is carried out at room temperature of 25°C. The specific process is: first, a constant current discharge is performed at a rate of 1C (corresponding to a current of 32A for a nominal battery capacity of 32Ah), and voltage, current and temperature data are collected in real time. When the voltage drops to the 2.5V protection threshold, the discharge is stopped and the battery is left to stand for 30 minutes; then, the battery is charged in a constant current mode of 32A to an upper limit voltage of 3.65V, and then switched to a constant voltage mode to continue charging until the cutoff current decays to 0.64A (0.02C). After charging is completed, the battery is left to stand for another 30 minutes to ensure stability; finally, a second discharge is performed at a constant current of 1C to a termination voltage of 2.5V, and the actual discharge capacity (Ah) of the single cell is recorded for subsequent capacity prediction. (12) Data cleaning and correction processing: During the battery cycle test, the battery parameters collected during the cycle test are prone to abnormal records due to environmental disturbances and equipment operation fluctuations. Such observation data that deviate from the actual physical characteristics need to be identified and corrected through statistical analysis methods. The processing flow includes two core links: first, based on Tukey's box plot method (confidence interval is set to 95%), outlier detection is performed on time series data such as voltage and current, and data outside the [Q1-1.5IQR, Q3+1.5IQR] interval is marked as outlier data; then, a sliding window mechanism (window width N=3) is used to perform median filtering on the abnormal points, specifically by interpolating the mean of its adjacent time series data to achieve data smoothing. This processing method effectively eliminates the measurement error caused by transient interference while retaining the distribution characteristics of the original data, ensuring the accuracy of subsequent battery capacity decay prediction. (13) Historical monitoring data of lithium batteries are collected through experiments, including parameters such as voltage, current, temperature, and the battery's state of charge and discharge (SOC). These lithium battery data are organized into a time series format to form an initial data set. The constructed data set can be expressed as In addition, the battery capacity needs to be determined as a label for the input data Therefore, the fully constructed dataset is (14) Based on the temporal correlation characteristics of battery cycle data, a moving window mechanism is used to construct model training samples. A fixed length L is set as the window size parameter. For the data segment corresponding to the cycle number c, all battery multidimensional features (including dynamic parameters such as voltage and current) and capacity measurement values ​​X in the interval [c, c+L] are extracted. input As the network input, the capacity truth value SOC of c+L+1 cycles is labelAs the prediction target. The window is moved backward along the time axis by a single-step sliding operation, and data samples with continuous time series association are generated in sequence until the end of the window reaches the end position of the data sequence. (15) In view of the dynamic range differences of battery terminal voltage, current and temperature parameters, the interval [-1,1] linear normalization method is introduced to achieve scale alignment of each feature channel by eliminating dimension differences. This preprocessing technology effectively suppresses gradient oscillation and promotes the stability of the neural network weight update process. Its mathematical definition is: Among them, x represents the original input data of the battery, x norm represents the standardized data, x max and x min Represents the maximum and minimum values ​​in the original input data. (16) The dataset is divided into training set, validation set, and test set. The ratio can be adjusted according to the actual situation. The division ratio is 70% training set, 15% validation set, and 15% test set.

[0099] More specifically, see Figure 2 In the above step 2 of the embodiment of the present invention, (11) the multi-scale feature extractor is designed based on the separation of large convolution kernels and a multi-path parallel topology architecture, including: utilizing a channel decoupling mechanism to perform convolution transformation on selected channels and retaining the remaining channels as identity mapping branches; specifically, it includes:

[0100] (111) By feeding the inputs of different branches into different parallel branches, the convolution operation of each branch can be expressed as:

[0101]

[0102] In formula (1), X′ hw Represents branch X hw Select the depth-separable convolution with convolution kernel size of 3×3 and 1×1 to extract effective features of the input data, X′ W Represents branch X w Select the depth-separable convolution with convolution kernel size of 1×7 and 1×1 to extract effective features of the input data, X′ h Represents branch X h Select the depth-separable convolution with convolution kernel size of 7×1 and 1×1 to extract effective features of the input data, X′ id Represents branch X idA depth-wise separable convolution with a convolution kernel size of 1×1 is used to extract effective features from the input data. The 1×1 convolution kernel is used to integrate channel information for feature combination and enhanced nonlinear mapping, the 3×3 convolution kernel is used for local spatial correlation, the 1×7 convolution kernel is used to extract long-range dependencies in the temporal dimension, and the 7×1 convolution kernel is used to extract cross-channel features. By rationally selecting and combining convolution kernels of different sizes, a balance is achieved between computational efficiency and multi-scale feature extraction capabilities, ensuring the efficiency and accuracy of the multi-scale feature extractor.

[0103] (112) Based on the separated large convolution kernel, the outputs of each branch are connected to obtain the connected output X′, which is expressed as:

[0104] X′=Concat(X′ hw ,X′ w ,X′ h ,X′ id ) Formula (2)

[0105] Regularization operation is performed on the connected output X′, and the expression is:

[0106] Y=Norm(X′) Formula (3)

[0107] In formula (3), Y represents the operation result;

[0108] The multi-scale feature extractor applies residual connections to improve gradient fluidity and network convergence speed during training, which can be expressed as:

[0109]

[0110] In formula (4), Y represents the operation result, r represents the expansion coefficient, C represents the input channel size, σ represents the activation function, and X represents the data set;

[0111] (12) The multi-scale feature extraction module in step 2 inputs the feature sequence X obtained in step 1 at the current moment input Processing, including:

[0112] The multi-scale feature extractor is used to extract the current input feature sequence X obtained in step 1. input , group them according to the channel dimension, and divide the channel dimension into two subsets, which are expressed as:

[0113] X hw ,X w ,X h ,X id =Split(X)

[0114] =X :,:g ,X :g:2g ,X :2g:3g ,X:3g: Formula (5)

[0115] In formula (5), X hw represents the extracted features related to height and width, X w represents the grouping feature assigned to the horizontal band kernel, X h represents the grouping feature assigned to the vertical band kernel, X id Indicates the extraction of identity or channel related features, X :,:g represents the first g columns of the input tensor X, X :g:2g Represents the gth to 2gth columns of the input tensor X, X :2g:3g represents the grouping feature assigned to the vertical band kernel, X :3g: Represents the 3gth to the last column of the input tensor X;

[0116] The multi-scale feature extractor is constructed by setting a ratio r n Determine the number of each branch channel, expressed as n = r n C, where n represents the number of channels of the convolution branch and C represents the number of channels of the input tensor;

[0117] (13) In step 2, the multi-path depth separation convolution operation is used to obtain the output features; wherein the output feature matrix X out1 Expressed as:

[0118] X out1 ∈R 4×P Formula (6)

[0119] In formula (6), R represents the expansion coefficient and P represents the dimension of the hidden layer;

[0120] Use the multi-scale feature extraction module to input the current moment into the feature sequence X input The output features are captured through multi-channel decomposition.

[0121] See also Figure 3 , the embodiment of the present invention applies depthwise separable convolution kernels of different sizes to perform depthwise separable convolution operations, so as to effectively extract and enrich multi-scale features, so that the multi-scale feature extractor can not only significantly expand the receptive field, but also better capture and understand long-distance dependencies, and effectively reduce the computational complexity through separation operations. The multi-scale feature extraction module in the embodiment of the present invention is based on the design of separate large convolution kernels and multi-way parallel topology architecture. It significantly improves the diversity and efficiency of feature extraction through channel decoupling mechanism and multi-path depthwise separable convolution operations. Among them, the channel decoupling mechanism is used to perform convolution transformation on the selected channels, and the remaining channels are retained as identity mapping branches, and a small 3×3 kernel is introduced as the basic convolution kernel. Taking into account the disadvantage of large kernels in computing speed, the large kernel k h ×kw Decomposed into 1×k w and k h ×1. This decomposition method not only improves the computational efficiency of the model, but also enhances the ability to extract multi-scale features, enabling more efficient feature learning and representation while maintaining model complexity. Compared with traditional methods, this design can more effectively capture multi-scale features and expand the receptive field, thereby providing richer and more accurate input features for the spatial relationship feature extractor. Specifically, the multi-scale feature extractor enhances the representation capability of node features, enabling the spatial relationship feature extractor to more accurately model the coupling relationship between battery parameters, thereby improving the accuracy of lithium battery SOC prediction.

[0122] In the above step 2 of the embodiment of the present invention, please refer to Figure 4 The features extracted by the graph convolutional network in the spatial relationship feature extractor of the embodiment of the present invention contain rich spatial information, reflect the current state of each node and the information fusion of adjacent nodes, and improve the representation ability of node features; including:

[0123] (21) Construct a feature interaction graph to describe the relationship between battery parameters, expressed as: G = (V, E), where G represents the feature interaction graph, V represents the node, and E represents the set of edges;

[0124] Each battery feature in the output feature is regarded as a node V, expressed as: V = {I, V, T, SOC} ∈ R 4×n , where I represents the unit matrix, T represents time, SOC represents the lithium battery SOC prediction, 4 represents the feature dimension of each time step, including voltage, current, temperature and capacity, and n represents the number of nodes in the historical data, where the feature dimension of each node at a specific moment reflects the state of the battery at that moment;

[0125] The physical correlation between features is represented by the adjacency matrix A, which is expressed as: A∈R 4×4 , A i,j ∈{0,1}, where A i,j Indicates whether there is an explicit coupling relationship between feature i and feature j;

[0126] Construct feature matrix X 4×P , select the current moment input feature sequence X in step 1 input After window division, it is used as the feature matrix X 4×P , the multi-dimensional time series monitoring data of the battery system is used as the node feature X, expressed as: X∈R 4×P , where each column represents a sampling moment, and each row contains the monitoring value sequence of the corresponding feature. If p = 1, then X t ∈R 4×1Represents the real-time feature vector at time t, including the current, voltage, temperature and capacity status at time t;

[0127] The current, voltage, temperature, and capacity states are constructed as nodes in a graph structure. The physical coupling relationship between parameters is captured through graph convolution operations to obtain a graph convolutional network, which enables interactive modeling of multi-dimensional spatial features.

[0128] The relationship between the feature interaction graph G and the adjacency matrix A includes: the adjacency matrix A is the mathematical representation of the feature interaction graph G, where whether there is an explicit coupling relationship A between feature i and feature j i,j That is, the physical connection strength between node i and node j. If A i,j >0, it means that there is an explicit coupling relationship between the nodes and they are connected by edge E; otherwise, there is no direct connection;

[0129] Please continue reading Figure 4 The detailed process of the spatial relationship feature extractor in the embodiment of the present invention to extract spatial information based on the feature interaction graph includes: taking the battery parameters as the node feature X, combining with the adjacency matrix A to construct a graph structure; aggregating the information of the node itself and its neighbors through multi-layer graph convolution operations, wherein the normalized adjacency matrix Ensure the stability of information transmission; each layer of graph convolution passes through the weight matrix W (l) The features are nonlinearly transformed and the enhanced spatial features H are output, which not only includes the node state itself, but also integrates the interactive information of adjacent nodes to improve the accuracy of SOC prediction.

[0130] Construct a feature interaction graph G, including: nodes V represent different parameters of the battery, including current, voltage, temperature and capacity state, edges E represent the physical coupling relationship between parameters, and feature matrix X 4×P Each row corresponds to a node, that is, each row corresponds to a parameter, and each row represents the monitoring value of the parameter at different time steps; the specific association is: the mapping between the node and the feature matrix, including the feature matrix X 4×P Each row of X i Represents the time series data of the i-th parameter, corresponding to node v in the figure i ; Among them, the first row: capacity sequence Q(t1),Q(t2),...,Q(t p ); The second row: voltage timing data V(t1), V(t2), ..., V(t p ); The third row: current time series data I(t1), I(t2), ..., I(t p ); The fourth row: temperature time series data T(t1), T(t2), ..., T(t p );

[0131] The physical meaning of the adjacency matrix A is: the adjacency matrix A defines the connection strength between nodes through prior knowledge. If the voltage and current have the Ohm's law relationship V = IR, then A 1,2 =1 (indicates strong coupling). If temperature affects internal resistance and thus voltage, then A 3,1 =1(temperature→voltage edge);

[0132] The graph convolution operation is performed through the adjacency matrix A and the feature matrix X 4×P The iterative aggregation of to achieve the modeling of the physical relationship between parameters is as follows:

[0133] Adjacency matrix normalization: add self-connection to A And normalize (to prevent gradient explosion):

[0134]

[0135] Where D represents the degree matrix,

[0136] Feature propagation and aggregation: Each layer of graph convolution operation updates the node features, expressed as:

[0137]

[0138] Node update: The new feature of a node is determined by a weighted combination of itself and its neighbors, expressed as:

[0139]

[0140] Among them, the weight w V , weight w I With weight w T By the learnable parameter W (l) Joint decision.

[0141] (22) The graph convolution operation corresponding to the graph convolution network is expressed as:

[0142]

[0143] In formula (7), X (l+1) Represents the node feature matrix of the l+1 layer, which represents the features of all nodes at the current level, X (l) represents the node feature matrix of the lth layer, σ represents the activation function, Represents the degree matrix, which is the degree of each node and is used to normalize the adjacency matrix. It means that the adjacency matrix A is increased by the identity matrix I, which is a self-connection between each node and itself, ensuring that each node can be updated with its own information. (l)Represents the learnable weight matrix of the lth layer, which is used to map node features to a new space. Its dimension is expressed as 4×P, where P represents the feature dimension of the hidden layer;

[0144] Specifically, the physical meaning of graph convolution operations is particularly important for battery SOC prediction. The relationship between voltage nodes and current nodes can be modeled using Ohm's law, which means that changes in current directly affect changes in battery voltage. Therefore, the voltage node can be fused with the current node through graph convolution operations. If the relationship between voltage nodes and current nodes is modeled using Ohm's law, meaning that changes in current directly affect changes in battery voltage, then the voltage node is fused with the current node through graph convolution operations. If the battery's temperature affects its internal resistance, which in turn affects voltage changes, reflecting the internal resistance temperature drift effect, the interaction between temperature and voltage is captured through graph convolution, allowing the voltage node to comprehensively consider the influence of temperature factors during calculation. If the transformation of the capacity node is significantly affected by temperature, especially during battery aging, where temperature changes accelerate the attenuation of battery capacity, the graph convolution network captures the battery aging effect through the relationship between the temperature node and the capacity node, facilitating accurate prediction of lithium battery SOC.

[0145] After multi-layer graph convolution operations, the final output feature matrix These are node features enhanced with spatial features, where L represents the number of graph convolution layers and P represents the dimension of the hidden layer. These enhanced features contain rich spatial information, not only reflecting the current state of each node but also integrating information from adjacent nodes, thereby improving the representation of node features. By extracting these spatial features, GCN provides a more accurate spatial feature representation for battery SOC prediction, helping to capture various physical phenomena such as battery aging, internal resistance changes, current-voltage relationships, and temperature effects, thereby improving the accuracy of SOC prediction.

[0146] Predicting the state of charge (SOC) of lithium batteries is a complex problem involving the coupling of multiple physical quantities. Its accuracy is highly dependent on modeling the dynamic relationships between parameters such as voltage, current, temperature, and capacity. The relationship between voltage and current follows the basic principle of Ohm's law, but the internal resistance of the battery varies with SOC, temperature, and aging, resulting in a nonlinear response. Graph convolutional networks construct a feature interaction graph to explicitly model the connections between voltage and current nodes, dynamically capturing voltage fluctuations caused by polarization effects during charging and discharging, thereby revising the SOC estimate. During high-current discharge, a sudden drop in voltage may indicate a rapid decrease in SOC. Traditional methods may ignore this dynamic coupling relationship, leading to error accumulation. Through adaptive adjustment of the adjacency matrix, the model can reflect the voltage-current characteristics under different operating conditions in real time, significantly improving the robustness of predictions in highly dynamic scenarios.

[0147] Temperature's impact on SOC prediction is primarily reflected in its dual effects on internal resistance and capacity. On the one hand, rising temperature reduces the battery's internal resistance, causing a change in the voltage response at the same current. If left uncorrected, this can lead to biased SOC estimation. In low-temperature environments, the increased internal resistance causes a significant voltage drop, which traditional methods may misinterpret as a decrease in SOC. However, this method automatically compensates for this drift by leveraging the temperature-voltage coupling relationship within a graph convolutional network. On the other hand, temperature has a long-term impact on battery capacity. High temperatures accelerate electrolyte decomposition and SEI film thickening, leading to irreversible capacity decay. The model leverages the interaction between temperature and capacity nodes and historical data to learn aging trends and dynamically adjust the capacity decay coefficient to avoid overestimating the actual available SOC due to battery aging. This mechanism ensures that the model maintains high accuracy throughout the battery lifecycle, making it particularly suitable for scenarios requiring reliable long-term predictions, such as energy storage systems and electric vehicles.

[0148] Furthermore, the aging process of lithium batteries gradually alters their internal characteristics, such as increased internal resistance and loss of active lithium ions, further complicating SOC prediction. Traditional methods typically rely on fixed battery models and struggle to adapt to parameter drift caused by aging. This method uses a multi-scale feature extractor to capture the slowly varying characteristics of internal resistance and integrates the voltage-current-temperature interaction with a spatial relationship feature extractor to achieve adaptive modeling of aging effects. A time series feature extractor analyzes the periodic characteristics and long-term trends of temperature fluctuations, thereby distinguishing between instantaneous temperature changes and capacity decay caused by aging. This multi-level, multi-physics fusion framework enables the model to simultaneously address short-term dynamic responses and long-term performance degradation, maintaining stable predictions under complex operating conditions. Finally, a weighted feature fusion module integrates spatial relationships, temporal dependencies, and multi-scale features to generate highly accurate SOC predictions, providing a reliable basis for state assessment in battery management systems.

[0149] (23) After multi-layer graph convolution operations, the final output is the same as X out1 Feature matrices of the same dimension is the node feature enhanced by spatial features, expressed as: Where L represents the number of graph convolution layers, and P represents the dimension of the hidden layer. The enhanced features contain rich spatial information, which not only reflects the current state of each node, but also integrates the information of adjacent nodes to improve the representation ability of node features. By extracting spatial features, the spatial relationship feature extractor provides a more accurate spatial feature representation for battery SOC prediction, which helps to capture various physical phenomena including the aging process of lithium batteries, internal resistance changes, current-voltage relationship, and temperature effects, thereby improving the accuracy of lithium battery SOC prediction.

[0150] See also Figure 5In the above step 2 of the embodiment of the present invention, the lithium battery SOC adaptive prediction network based on the multi-scale graph convolutional network constructed further includes at least: a time series feature extractor and a feature weighted fusion module, specifically:

[0151] The time series feature extractor uses a multi-scale multi-head self-attention mechanism to capture the sequence correlation within each time scale, and uses dimension redefinition to reshape the tensor of each scale into a two-dimensional matrix. It then aggregates different scales according to the magnitude of the two-dimensional matrix, effectively fusing the time series features into the feature weighted fusion module, and dynamically weighting and integrating them with spatial relationship features and time series features.

[0152] The feature fusion weighting module integrates multi-scale features, spatial relationship features and time series features into a unified framework, and fully integrates them through the weight parameters of the scientific department to obtain the fusion features;

[0153] Preferably, during the training process of the lithium battery SOC adaptive prediction network based on the multi-scale graph convolutional network, the fusion features are input into the prediction module to predict the SOC of the lithium battery, and the various parameters of the lithium battery SOC adaptive prediction network are adjusted to obtain the lithium battery SOC prediction result.

[0154] The time series feature extractor of the embodiment of the present invention uses fast Fourier transform to detect the input feature sequence X at the current moment. input The periodic characteristics and assigning an appropriate time scale to each period are expressed as:

[0155]

[0156] In formula (8), FFT(·) represents the input feature sequence X at the current moment input For FFT calculation, Amp(·) represents the calculation of amplitude, F represents a vector, and the amplitude of each frequency reflects the periodic characteristics on different time scales, f1,…,f k Represents the frequency amplitude vector, whose dimension is the same as the number of frequency components after Fourier transform, and each element represents the amplitude value of the corresponding frequency, s i Represents the frequency index, which is used to traverse the frequency components after Fourier transform, f i represents the i-th frequency amplitude vector, and L represents the number of graph convolution layers;

[0157] Based on the selected time scale {s1,…,s k}, the input data is reshaped into a 3D tensor by the following formula to obtain several representations corresponding to different time scales, the expression is:

[0158]

[0159] In formula (9), Padding(·) is used to extend the time series by zero along the time dimension, X (l) represents the reshaped feature matrix, represents the i-th reshaped time series based on time scale i, Represents the dimension reshaping function, the first dimension of the reshaped tensor is s i , the second dimension is f i ;

[0160] For the feature tensor F after Fourier transform processing, the multi-head self-attention mechanism is applied on its time scale dimension to extract the time correlation features, which can be expressed as:

[0161]

[0162] In formula (10), Represents the tensor of each scale, MHA S (·) represents the multi-head attention function in the scale dimension;

[0163] The tensor of each scale Shaped into a two-dimensional matrix The tensors are weighted and aggregated according to the amplitude of each scale, and the frequency values ​​in the Fourier transform results are Apply the SoftMax function to calculate the amplitude, the expression is:

[0164]

[0165] In formula (11), represents the output of the SoftMax function, represents the normalized weights of k branches, represents the final weighted fusion output, represents the i-th normalized weight, represents the independent output of k branches;

[0166] Features at different scales are weighted and aggregated according to their amplitude, emphasizing the significant information at each scale. By calculating the amplitude, the time series feature extractor is helped to adaptively focus on the most relevant temporal features based on the periodicity strength of each scale. The expert mixture strategy enables the time series feature extractor to efficiently integrate multi-scale time series data and extract more comprehensive and detailed time-dependent features.

[0167] Output time correlation feature X out3 , denoted as X out3 ∈R Px4 , in order to enhance the performance of time series feature extractor in multi-scale time series modeling, capture the temporal correlation at different scales, and better cope with lithium battery SOC prediction.

[0168] The feature fusion module in the embodiment of the present invention fuses the multi-scale features extracted by the multi-scale feature extractor, the spatial relationship features extracted by the spatial relationship feature extractor, and the time series features extracted by the time series feature extractor during the information fusion stage, and outputs the fused features using a fully connected layer; specifically:

[0169] It is preferred to adopt a dynamic weight learning mechanism to further improve the flexibility of fusion, introduce learnable attention weights to adapt to the feature changes of different battery charging and discharging stages, and automatically adjust the fusion ratio of multi-scale, spatial and temporal features according to the distribution of input data.

[0170] The multi-scale features, spatial relationship features and time series features are fused using the weighted summation method. Assume that the feature vectors of the three scales are the multi-scale features X out1 , spatial relationship feature X out2 and time series features X out3 , whose corresponding weight coefficients are w1, w2, w3 respectively, and the feature vector X in the fusion feature after the feature weighted fusion module is fused out Expressed as:

[0171] X out =w1·X out1 +w2·X out2 +w3·X out3 Formula (12)

[0172] The first weight coefficient w1, the second weight coefficient w2, and the third weight coefficient w3 are obtained through learning, and the contribution ratio of each scale feature is automatically adjusted according to the training process to achieve more flexible feature fusion.

[0173] In the above step 2 of the embodiment of the present invention, in the prediction stage, the fused features are input into a fully connected layer in the prediction module. The fully connected layer adjusts its output through an activation function to generate the final prediction result. The expression of the fully connected layer is:

[0174] Y=Sigmoid(Linear(W·X out +b)) Formula (12)

[0175] In formula (12), Sigmoid represents the activation function, which is used to increase nonlinearity and improve the expressiveness of the model, W represents the weight matrix of the fully connected layer, X out represents the fused feature vector, and b represents the bias vector of the fully connected layer.

[0176] In the above step 2 of the embodiment of the present invention, the lithium battery SOC adaptive prediction network based on the multi-scale graph convolutional network is trained, including:

[0177] (11) Construct and preprocess dataset D t , and for this data set D t Divide into training set, validation set and test set; among them, the data set D t Expressed as: Specifically, the training data set constructed in step 1 is input into the SOC prediction network for training, and the verification data set is also input into the network for verification. The loss function of the model adopts the mean square error loss function, and the Adam optimizer is used to update and adjust the model parameters to reduce the prediction error.

[0178] (12) The lithium battery SOC adaptive prediction network is trained and verified using the training set and the test set. The optimizer is used to update and adjust the model parameters, and its performance on the validation set is continuously monitored during the training process. The parameters are adjusted to prevent overfitting, and a trained SOC adaptive prediction network is obtained. Specifically, the embodiment of the present invention continuously monitors the performance of the network on the validation data set during the training process to adjust hyperparameters and prevent overfitting. After the training is completed, the test data set is input into the network for final testing, and the prediction performance is evaluated using indicators such as the mean absolute error (MAE) and the root mean square error (RMSE). RMSE is used to measure the stability of the predicted SOC, which can more accurately reflect the degree of fluctuation between the predicted value and the true value. MAE is used to quantify the absolute deviation between the predicted SOC and the actual SOC, providing an intuitive description of the error amplitude. If the prediction result of the model on the test set does not achieve the expected effect, it is necessary to optimize the model training process by adjusting parameters such as the network structure, learning rate, and regularization method until satisfactory results are obtained. Finally, the optimized network architecture and trained model parameters are saved to facilitate subsequent applications and the execution of reasoning tasks, which is expressed as:

[0179]

[0180] Where y i represents the true value, represents the predicted value, and N represents the total number of samples.

[0181] (13) After the training is completed, the test set is input into the trained lithium battery SOC adaptive prediction network for testing, and its prediction performance and stability are evaluated to accurately predict SOC.

[0182] The experimental data of the embodiment of the present invention comes from a laboratory lithium-ion battery life test experiment. The rated capacity of the battery is 32Ah. The number of battery cycles and the battery capacity of the lithium-ion battery in each cycle are extracted and stored through the experiment. The charge and discharge data group contains time, charge and discharge voltage, and charge and discharge current data structure, including the number of battery cycles and battery capacity, etc. First, a constant current discharge is performed at a rate of 1C (corresponding to a current of 32A with a nominal capacity of 32Ah of the battery), and voltage, current and temperature data are collected in real time. When the voltage drops to the 2.5V protection threshold, the discharge is stopped and allowed to stand for 30 minutes; then, the battery is charged in a 32A constant current mode to an upper limit voltage of 3.65V, and then switched to a constant voltage mode for continuous charging until the cutoff current decays to 0.64A (0.02C). After charging is completed, the battery is allowed to stand for 30 minutes again to ensure that the state is stable; finally, the battery is discharged twice at a constant current of 1C to a termination voltage of 2.5V, and the actual discharge capacity value of the single battery is recorded for subsequent capacity prediction.

[0183] During data preprocessing, a boxplot outlier detection method is used to detect outliers in time series data such as voltage and current. Data outside the interval [Q1-1.5IQR, Q3+1.5IQR] is marked as outliers, and abnormal data points outside the range of three standard deviations are removed and smoothed. Furthermore, this embodiment of the present invention utilizes a sliding window to construct model training samples. The standardized battery degradation data is imported into a multi-scale graph convolutional framework to construct an adaptive prediction model for lithium-ion battery state of charge based on dynamic feature perception. The network parameters are trained through a supervised learning mechanism.

[0184] See also Figure 6 、 Figure 7 and Figure 8 The network structure and parameter settings built by the embodiment of the present invention are as follows Figure 6 The simulation experiment result evaluation table is shown in Figure 7 As shown in the figure, the original data curve and prediction curve after network training are as follows: Figure 8 As shown, the embodiment of the present invention constructs a graphical representation of battery parameters and uses a graph convolutional network to capture the spatial dependencies between parameters. In order to capture the multi-scale dynamic changes in the battery charging and discharging process, the multi-scale characteristics of the lithium battery are comprehensively considered. By combining the multi-scale analysis technology, the features at different time scales are extracted to better capture the dynamic changes in the battery charging and discharging process, effectively improve the accuracy and robustness of SOC prediction, improve the performance of the lithium battery management system and extend the battery life.

[0185] The above description is only a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with this technical field can easily think of various equivalent modifications or replacements within the technical scope disclosed in the present invention, and these modifications or replacements should all be included in the scope of protection of the present invention.

Claims

1. A lithium battery SOC adaptive prediction method based on a multi-scale graph convolutional network, characterized in that: include: Step 1: Obtain the monitoring data of the lithium battery and perform data cleaning and standardization to obtain the current input feature sequence X input ; Step 2: Input the current moment into the feature sequence X input Input to the lithium battery SOC adaptive prediction network based on the multi-scale graph convolutional network, and output the lithium battery SOC prediction result; wherein: the constructed lithium battery SOC adaptive prediction network based on the multi-scale graph convolutional network includes: a multi-scale feature extractor, a spatial relationship feature extractor and a prediction module, specifically: The multi-scale feature extractor is designed based on the separation of large convolution kernels and multi-path parallel topology architecture. The multi-scale feature extractor extracts the current moment input feature sequence X obtained in step 1 through the multi-scale feature extraction module. input Processing is performed and multi-path depth separation convolution operation is used to obtain output features; The features extracted by the graph convolutional network in the spatial relationship feature extractor contain rich spatial information, reflecting the current state of each node and the information fusion of adjacent nodes, thereby improving the representation ability of node features; The prediction module predicts the SOC of the lithium battery to obtain a lithium battery SOC prediction result.

2. The lithium battery SOC adaptive prediction method based on a multi-scale graph convolutional network according to claim 1 is characterized in that: In the step 2, (11) the multi-scale feature extractor is designed based on a separated large convolution kernel and a multi-path parallel topology architecture, including: utilizing a channel decoupling mechanism to perform a convolution transformation on the selected channels and retaining the remaining channels as identity mapping branches; specifically, including: (111) By feeding the inputs of different branches into different parallel branches, the convolution operation of each branch can be expressed as: In formula (1), X′ hw Represents branch X hw Select the depth-separable convolution with convolution kernel size of 3×3 and 1×1 to extract effective features of the input data, X′ W Represents branch X w Select the depth-separable convolution with convolution kernel size of 1×7 and 1×1 to extract effective features of the input data, X′ h Represents branch X h The depth-wise separable convolution with kernel size of 7×1 and 1×1 is used to extract effective features from the input data. i ' d Represents branch X id A depth-wise separable convolution with a convolution kernel size of 1×1 is used to extract effective features from the input data. The 1×1 convolution kernel is used to integrate channel information for feature combination and enhanced nonlinear mapping, the 3×3 convolution kernel is used for local spatial correlation, the 1×7 convolution kernel is used to extract long-range dependencies in the time dimension, and the 7×1 convolution kernel is used to extract cross-channel features. (112) Based on the separated large convolution kernel, the outputs of each branch are connected to obtain the connected output X′, which is expressed as: X′ = Concat(X′ hw , X′ w , X′ h , X′ id ) Equation (2) Regularization operation is performed on the connected output X′, and the expression is: Y=Norm(X′) Formula (3) In formula (3), Y represents the operation result; The multi-scale feature extractor uses residual connections to improve gradient fluidity and network convergence speed during training, which can be expressed as: In formula (4), Y represents the operation result, r represents the expansion coefficient, C represents the input channel size, σ represents the activation function, and X represents the data set; (12) The multi-scale feature extraction module in step 2 extracts the current moment input feature sequence X obtained in step 1. input Processing, including: The multi-scale feature extractor is used to extract the current input feature sequence X obtained in step 1. input , group them according to the channel dimension, and divide the channel dimension into two subsets, which are expressed as: X hw , X w , X h , X id = Split(X) = X :,:g , X :g:2g , X :2g:3g , X :3g: Equation (5) In formula (5), X hw represents the extracted features related to height and width, X w represents the grouping feature assigned to the horizontal band kernel, X h represents the grouping feature assigned to the vertical band kernel, X id Indicates the extraction of identity or channel related features, X :,:g represents the first g columns of the input tensor X, X :g:2g Represents the gth to 2gth columns of the input tensor X, X :2g:3g represents the grouping feature assigned to the vertical band kernel, X :3g: Represents the 3gth to the last column of the input tensor X; The multi-scale feature extractor is constructed by setting a ratio r n Determine the number of each branch channel, expressed as n = r n C, where n represents the number of channels of the convolution branch and C represents the number of channels of the input tensor; (13) In step 2, a multi-path depth separation convolution operation is used to obtain output features; wherein the output feature matrix X contained in the output feature is out1 Expressed as: X out1 ∈R 4×P Formula (6) In formula (6), R represents the expansion coefficient and P represents the dimension of the hidden layer; The multi-scale feature extraction module is used to input the feature sequence X at the current moment input The output features are captured through multi-channel decomposition.

3. The lithium battery SOC adaptive prediction method based on a multi-scale graph convolutional network according to claim 1 is characterized in that: The features extracted by the graph convolutional network in the spatial relationship feature extractor in step 2 contain rich spatial information, reflecting the current state of each node and the information fusion of adjacent nodes, thereby improving the representation capability of node features; include: (21) Construct a feature interaction graph to describe the relationship between battery parameters, expressed as: G = (V, E), where G represents the feature interaction graph, V represents the node, and E represents the set of edges; Each battery feature in the output features is regarded as a node V, expressed as: V = {I, V, T, SOC} ∈ R 4×n , where I represents the identity matrix, T represents time, SOC represents the lithium battery SOC prediction, and 4 represents the feature dimension of each time step; The physical correlation between features is represented by the adjacency matrix A, which is expressed as: A∈R 4×4 , A i,j ∈{0,1}, where A i,j Indicates whether there is an explicit coupling relationship between feature i and feature j; Construct feature matrix X 4×P , select the current moment input feature sequence X in step 1 input After window division, it is used as the feature matrix X 4×P , the multi-dimensional time series monitoring data of the battery system is used as the node feature X, expressed as: X∈R 4×P ; The current, voltage, temperature, and capacity states are constructed as nodes in a graph structure. The physical coupling relationship between parameters is captured through graph convolution operations to obtain a graph convolutional network, which enables interactive modeling of multi-dimensional spatial features. (22) The graph convolution operation corresponding to the graph convolution network is expressed as: In formula (7), X (l+1) Represents the node feature matrix of the l+1 layer, which represents the features of all nodes at the current level, X (l) represents the node feature matrix of the lth layer, σ represents the activation function, represents the degree matrix, Indicates that the adjacency matrix A is increased by the identity matrix I, W (l) represents the learnable weight matrix of layer l; (23) After multi-layer graph convolution operations, the final output is the same as X out1 Feature matrices of the same dimension is the node feature enhanced by spatial features, expressed as: Where L represents the number of graph convolution layers and P represents the dimension of the hidden layer.

4. The lithium battery SOC adaptive prediction method based on a multi-scale graph convolutional network according to claim 1 is characterized in that: In the step 2, the lithium battery SOC adaptive prediction network based on the multi-scale graph convolutional network constructed further includes at least: a time series feature extractor and a feature weighted fusion module, specifically: The time series feature extractor uses a multi-scale multi-head self-attention mechanism to capture the sequence correlation within each time scale, and uses dimension redefinition to reshape the tensor of each scale into a two-dimensional matrix; The feature fusion weighting module integrates multi-scale features, spatial relationship features and time series features into a unified framework, and fully integrates them through the weight parameters of the scientific system to obtain fusion features; Preferably, during the training process of the lithium battery SOC adaptive prediction network based on the multi-scale graph convolutional network, the fusion features are input into the prediction module to predict the SOC of the lithium battery, and the various parameters of the lithium battery SOC adaptive prediction network are adjusted to obtain the lithium battery SOC prediction result.

5. The lithium battery SOC adaptive prediction method based on a multi-scale graph convolutional network according to claim 4 is characterized in that: The time series feature extractor uses fast Fourier transform to detect the current input feature sequence X input The periodic characteristics and assigning an appropriate time scale to each period are expressed as: In formula (8), FFT(·) represents the input feature sequence X at the current moment input For FFT calculation, Amp(·) represents the calculation of amplitude, F represents a vector, and the amplitude of each frequency reflects the periodic characteristics on different time scales, f1,…,f k Represents the frequency amplitude vector, whose dimension is the same as the number of frequency components after Fourier transform, and each element represents the amplitude value of the corresponding frequency, s i Represents the frequency index, which is used to traverse the frequency components after Fourier transform, f i represents the i-th frequency amplitude vector, and L represents the number of graph convolution layers; Based on the selected time scale {s1,…,s k }, the input data is reshaped into a 3D tensor by the following formula to obtain several representations corresponding to different time scales, the expression is: In formula (9), Padding(·) is used to extend the time series by zero along the time dimension, X (l) represents the reshaped feature matrix, represents the i-th reshaped time series based on time scale i, Represents the dimension reshaping function, the first dimension of the reshaped tensor is s i , the second dimension is f i ; For the feature tensor F after Fourier transform processing, the multi-head self-attention mechanism is applied on its time scale dimension to extract the time correlation features, which can be expressed as: In formula (10), Represents the tensor of each scale, MHA S (·) represents the multi-head attention function in the scale dimension; The tensor of each scale Shaped into a two-dimensional matrix The tensors are weighted and aggregated according to the amplitude of each scale, and the frequency values ​​in the Fourier transform results are Apply the SoftMax function to calculate the amplitude, the expression is: In formula (11), represents the output of the SoftMax function, represents the normalized weights of k branches, represents the final weighted fusion output, represents the i-th normalized weight, represents the independent output of k branches; Features at different scales are weighted and aggregated according to their amplitudes, emphasizing significant information at each scale. By calculating the amplitude, the time series feature extractor is helped to adaptively focus on the most relevant temporal features based on the periodicity strength of each scale. Based on the expert mixture strategy, the time series feature extractor is enabled to efficiently integrate multi-scale time series data and extract more comprehensive and detailed time-dependent features. Output time correlation feature X out3 , denoted as X out3 ∈R Px4 , in order to enhance the performance of the time series feature extractor in multi-scale time series modeling, capture the temporal correlation at different scales, and better cope with lithium battery SOC prediction.

6. The method for adaptively predicting lithium battery SOC based on a multi-scale graph convolutional network according to claim 4, characterized in that: The feature fusion module fuses the multi-scale features extracted by the multi-scale feature extractor, the spatial relationship features extracted by the spatial relationship feature extractor, and the time series features extracted by the time series feature extractor in the information fusion stage, and outputs the fused features using a fully connected layer; specifically: The dynamic weight learning mechanism further enhances the flexibility of fusion. It introduces learnable attention weights to adapt to the characteristic changes of different battery charging and discharging stages, and automatically adjusts the fusion ratio of multi-scale, spatial, and temporal features according to the distribution of input data. The multi-scale features, spatial relationship features and time series features are fused using the weighted summation method. Assume that the feature vectors of the three scales are the multi-scale features X out1 , spatial relationship feature X out2 and time series features X out3 , whose corresponding weight coefficients are w1, w2, w3 respectively, and the feature vector X in the fusion feature after fusion by the feature weighted fusion module out Expressed as: X out =w1·X out1 +w2·X out2 +w3·X out3 Formula (12) The first weight coefficient w1, the second weight coefficient w2, and the third weight coefficient w3 are obtained through learning, and the contribution ratio of each scale feature is automatically adjusted according to the training process to achieve more flexible feature fusion.

7. The method for adaptively predicting lithium battery SOC based on a multi-scale graph convolutional network according to claim 4, characterized in that: In the prediction stage, the fusion features are input into a fully connected layer in the prediction module. The fully connected layer adjusts its output through the activation function to generate the final prediction result. The expression of the fully connected layer is: Y = Sigmoid(Linear(W·X out + b)) Equation (12) In formula (12), Sigmoid represents the activation function, which is used to increase nonlinearity and improve the expressiveness of the model, W represents the weight matrix of the fully connected layer, X out represents the fused feature vector, and b represents the bias vector of the fully connected layer.

8. The method for adaptively predicting lithium battery SOC based on a multi-scale graph convolutional network according to claim 4, characterized in that: Training of the lithium battery SOC adaptive prediction network based on a multi-scale graph convolutional network, including: Construct and preprocess dataset D t , and for this data set D t Divide into training set, validation set and test set; among them, the data set D t Expressed as: The lithium battery SOC adaptive prediction network is trained and optimized using the training set and test set. The optimizer is used to update and adjust the model parameters. During the training process, the performance on the validation set is continuously monitored and the parameters are adjusted to prevent overfitting. The trained SOC adaptive prediction network is obtained. After training, the test set is input into the trained lithium battery SOC adaptive prediction network for testing, and its prediction performance and stability are evaluated to accurately predict SOC.

Images