A lithium battery capacity prediction method based on dual-path feature fusion and spatiotemporal modeling

Abstract

This invention provides a lithium battery capacity prediction method based on dual-path feature fusion and spatiotemporal modeling, belonging to the field of battery state prediction technology in computer data processing. It collects time-series data of lithium battery health factors, and based on a lithium battery operating state feature extraction model, obtains the intrinsic mode function components of lithium battery health factors through principal component analysis and designed intrinsic mode function decomposition. Through two feature extraction paths of a deep residual module, it obtains graph features and projection features of the lithium battery health factors. After enhancing two types of features through graph networks and multilayer perceptrons, it obtains a causal feature enhancement network to obtain a capacity causal feature vector. This vector is input into the prediction module to output the predicted capacity value. This invention overcomes the shortcomings of existing technologies in feature extraction, fusion, and time-series modeling, improving the accuracy and robustness of capacity prediction.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of battery state prediction in computer data processing, and particularly relates to a lithium battery capacity prediction method based on double-path feature fusion and space-time modeling. BACKGROUND

[0002] Lithium batteries are widely used in electric vehicles, energy storage systems and portable electronic devices due to their high energy density, long cycle life and environmental protection characteristics. The state of health (SOH) of lithium batteries directly affects their performance and safety, and accurate prediction of the capacity is a key indicator for evaluating SOH. However, the capacity decay process of lithium batteries has characteristics such as nonlinearity, time-varying and multi-factor coupling, and traditional prediction methods are difficult to achieve high precision and strong generalization ability.

[0003] In the prior art, data-driven lithium battery capacity prediction methods include:

[0004] A Chinese invention patent with application number CN202510670207.8 discloses a lithium battery life prediction method combining multi-level feature fusion and time series modeling, including the following steps: extracting local feature representation of battery capacity time series data through one-dimensional convolutional neural network; using the Transformer multi-head self-attention structure to perform multi-scale global fusion on the local features to obtain global fusion features; using a bidirectional long short-term memory network structure to capture dynamic evolution features of the battery operating state in the forward and reverse time dimensions; finally outputting the remaining useful life prediction value of the battery through a fully connected regression layer. A Chinese invention patent with application number CN202510266372.7 discloses a battery RUL prediction method based on variable mode decomposition, including the following steps: obtaining battery capacity degradation data and battery capacity failure threshold; using variable mode decomposition to smooth the battery capacity degradation data to obtain multiple modal components; obtaining battery capacity key feature data according to the multiple modal components; inputting the battery capacity key feature data into an optimized support vector regression model to obtain a battery capacity prediction value; and when the battery capacity prediction value is less than the battery capacity failure threshold, calculating the remaining useful life of the battery.

[0005] The above methods and prior art have the following limitations:

[0006] 1) The capacity decay of lithium batteries is affected by multiple factors, and the above technology fails to effectively fuse multi-dimensional health factors, resulting in incomplete feature extraction, and the modal decomposition parameters are fixed and cannot be adaptively optimized, which may ignore key modal components and affect prediction accuracy;

[0007] 2) The above technical feature extraction path is single, lacks a multi-scale feature fusion mechanism, is difficult to capture a space-time dependent relationship, and has poor generalization ability without combining a graph attention space model.

[0008] The present application aims to propose a lithium battery capacity prediction method based on dual-path feature fusion and space-time modeling to solve the deficiencies of the prior art in feature extraction, fusion and time series modeling, and improve the accuracy and robustness of capacity prediction. SUMMARY

[0009] To solve the above problems, the present application aims to propose a lithium battery capacity prediction method based on dual-path feature fusion and space-time modeling to solve the deficiencies of the prior art in feature extraction, fusion and time series modeling, and improve the accuracy and robustness of capacity prediction.

[0010] The present application proposes a lithium battery capacity prediction method based on dual-path feature fusion and space-time modeling, comprising the following steps:

[0011] S1, collecting a plurality of lithium battery health factor time series data affecting the capacity in a time window;

[0012] S2, inputting the collected data into a lithium battery operating state feature extraction model, the model comprising a lithium battery health factor feature decomposition module and a deep residual module; the lithium battery health factor feature decomposition module comprises principal component analysis and intrinsic mode function decomposition, and obtains lithium battery health factor intrinsic mode function components; the deep residual module comprises two feature extraction paths, the first path performs time series coding and spatial feature extraction on the lithium battery health factor time series data to obtain lithium battery health factor graph features; the second path projects and dimensionally matches the lithium battery health factor intrinsic mode function components to obtain intrinsic mode function projection features;

[0013] S3, enhancing the features obtained in S2 based on a lithium battery capacity causal feature enhancement module; inputting the lithium battery health factor graph features into a graph structure nonlinear enhancement network to obtain enhanced lithium battery health factor graph features; inputting the intrinsic mode function projection features into a multilayer perception machine to obtain enhanced intrinsic mode function projection features; inputting the two enhanced features into a causal feature enhancement network to obtain a lithium battery capacity causal feature vector;

[0014] S4, inputting the lithium battery capacity causal feature vector into a lithium battery capacity prediction module to obtain a lithium battery capacity prediction value.

[0015] Preferably, the lithium battery health factor time series data comprises:

[0016] The highest temperature and the lowest temperature values of all probes at each time in the time window form the highest temperature time series data and the lowest temperature time series data;

[0017] The maximum value and the minimum value of the single cell voltage at each time in the time window are subtracted to obtain the instantaneous voltage difference, and then the highest voltage time series data, the lowest voltage time series data and the voltage difference time series data are composed.

[0018] The current time series and the battery pack voltage time series on the main loop of the lithium battery module in the time window.

[0019] Preferably, the lithium battery health factor feature decomposition module specifically comprises:

[0020] First, the obtained lithium battery health factor time series data is subjected to principal component analysis to obtain a reduced dimension health factor matrix;

[0021] Then, an intrinsic mode function decomposition model based on multi-objective optimization is constructed to perform variational modal decomposition on the reduced dimension health factor matrix, including:

[0022] S21, a lithium battery health factor intrinsic mode function decomposition multi-objective optimization problem is constructed; the lithium battery health factor intrinsic mode function decomposition multi-objective optimization problem includes a multi-objective weighted objective function and a decision variable; the multi-objective weighted objective function is calculated by weighted summation of four objective functions; the first objective function is used to evaluate the correlation between the intrinsic mode function component and the original signal, which is quantified by a weighted correlation; the second objective function is used to evaluate the proportion of effective intrinsic mode function, which is quantified by the proportion of effective intrinsic mode function; the third objective function is used to measure the accuracy of signal reconstruction after variational modal decomposition, which is quantified by reconstruction error; the fourth objective function reflects the average complexity of the intrinsic mode function component, which is quantified by average information entropy; the decision variable is the mode number in the variational modal decomposition And the penalty factor ;

[0023] S22, a high-dimensional multi-objective particle swarm optimization algorithm based on reference vector guidance and inflection point optimization strategy is used to solve the lithium battery health factor intrinsic mode function decomposition multi-objective optimization problem; a set of predefined reference vectors are introduced into the traditional particle swarm algorithm to guide the uniform distribution of the particle swarm in the four-dimensional space, and the angle penalty distance is used for environmental selection, and finally the optimal mode number And the optimal penalty factor ; the optimal mode number and the optimal penalty factor are used to configure the variational modal decomposition method to perform variational modal decomposition on the lithium battery health factor, and the lithium battery health factor intrinsic mode function component is obtained.

[0024] Preferably, the data processing process of the first path of the deep residual module is specifically:

[0025] Firstly, a time series encoder is constructed to encode the lithium battery health factors in time series; the encoder includes an input layer, a regularization layer, a ReLU activation function layer and a random inactivation layer; the lithium battery health factors are input into the encoder for dimension reduction to obtain the time series encoding features of the lithium battery health factors;

[0026] Then, a lithium battery health factor space feature extraction module is constructed:

[0027] A lithium battery health factor full connection graph is constructed , wherein , ; wherein the health factors are taken as nodes , the health factor data is taken as node features, and the interaction between the health factors is taken as edges connecting two health factors , and the edge length represents the influence size between the health factors, and all edge lengths are set to 1;

[0028] A lithium battery health factor space feature extraction network is constructed to calculate the weight of each edge in the lithium battery health factor full connection graph and extract the health factor space features; the network includes a multi-head graph attention convolution layer, a ReLU activation function layer, a random inactivation layer, a single-head graph attention convolution layer and a graph aggregation layer; firstly, the node features in the lithium battery health factor full connection graph are input into the multi-head graph attention convolution layer to obtain the multi-head space attention features of each node in the full connection graph, and the features of all nodes constitute the lithium battery health factor multi-head space attention features; secondly, the lithium battery health factor multi-head space attention features are input into the ReLU activation function layer for feature activation; thirdly, the features output by the ReLU activation function layer are input into the random inactivation layer to suppress the expression of part of the neurons in the model, and the lithium battery health factor space advanced features are obtained; the lithium battery health factor space advanced features are input into the single-head graph attention convolution layer to perform attention coefficient calculation and channel feature value weighting on each feature channel, and the lithium battery health factor space deep features are obtained; the lithium battery health factor space deep features are input into the regularization layer for layer normalization operation on the feature values; after the normalized lithium battery health factor space deep features are input into the random inactivation layer, they are input into the graph aggregation layer to calculate the mean value of each feature channel and update the features of each node in the lithium battery health factor full connection graph, and the lithium battery health factor graph features are obtained.

[0029] Preferably, the data processing process of the second path of the deep residual module is specifically:

[0030] The intrinsic modal function component obtained is input into the input layer of the time sequence mapping network for linear projection, the feature dimension is mapped to 64 dimensions, and the normalized layer is input for standardization processing; the normalized features are input into the ReLU activation function layer for nonlinear activation, then input into the random inactivation layer to randomly suppress the expression of part of the neurons, to obtain the intrinsic modal function component intermediate features; secondly, the intrinsic modal function component intermediate features are input into the output layer of the multilayer perception to perform secondary dimension projection, and the feature dimension is mapped to 32 dimensions to match the lithium battery health factor graph feature obtained by the first path; the mapped features are input into the normalization layer to perform layer normalization operation on the feature values, to obtain the intrinsic modal function projection features.

[0031] Preferably, the lithium battery capacity causal feature enhancement module specifically comprises:

[0032] Firstly, the lithium battery health factor graph features are input into the graph structure nonlinear enhancement network, and the intrinsic modal function projection features are input into the multilayer perception, for nonlinear mapping and feature enhancement, to obtain enhanced lithium battery health factor graph features and intrinsic modal function features; wherein the graph structure nonlinear enhancement network is stacked by a graph convolution layer and a ReLU activation function layer, and each graph convolution layer is followed by a ReLU activation function layer; the multilayer perception is composed of a single full connection layer group layer;

[0033] Secondly, the enhanced lithium battery health factor graph features and the intrinsic modal function projection features are spliced in the feature dimension to form a joint feature vector; the joint feature vector is input into an attention layer containing a Tanh activation function to calculate a correlation coefficient; the correlation coefficient is normalized into a probability distribution by a Softmax function to generate an attention weight vector containing two feature weight; finally, a weighted feature fusion operation is performed, and the attention weight vector calculated is used to perform element-wise weighted multiplication on the enhanced lithium battery health factor graph features and the intrinsic modal function projection features, respectively, to obtain weighted lithium battery health factor graph features and weighted intrinsic modal function projection features; the two features after weighting are spliced again to be fused into a lithium battery capacity causal feature vector.

[0034] Preferably, the lithium battery capacity prediction module comprises a cascaded double-layer residual network, a long short-term memory network and a lithium battery capacity decoding network; the lithium battery capacity causal feature vector is input into the double-layer residual network to obtain enhanced lithium battery capacity causal features; the enhanced lithium battery capacity causal features are input into the long short-term memory network to obtain global time sequence features; the global time sequence features are input into the lithium battery capacity decoding network to obtain a lithium battery capacity prediction value.

[0035] Preferably, the cascaded two-layer residual network comprises two residual blocks for deep feature extraction of the obtained lithium battery capacity causal feature vector. The calculation process of each residual block is as follows: First, the input features are sequentially passed through a linear transformation layer, a normalization layer, and a ReLU activation function layer for feature mapping and nonlinear activation. Then, through a skip connection mechanism, the original input features of the residual block are directly superimposed onto the transformed main path features. If the input and output dimensions do not match, the input features are linearly projected to align the dimensions. The output of the first residual block is used as the input of the second residual block, and the above operation process is repeated to finally obtain the lithium battery capacity causal features enhanced by deep residuals.

[0036] Preferably, the lithium battery capacity decoding network consists of four layers of a multilayer perceptron, which maps the global temporal features output by the long short-term memory network to predicted lithium battery capacity values. Specifically, the first multilayer perceptron includes a linear input layer, a normalization layer, a ReLU activation function layer, and a random deactivation layer. The linear input layer performs a linear mapping on the global temporal features to obtain higher-dimensional features richer in lithium battery degradation characteristics. The higher-dimensional features are then sequentially input into the normalization layer, the ReLU activation function layer, and the random deactivation layer, where feature value normalization, nonlinear activation, and random neuron deactivation operations are performed respectively to obtain... High-dimensional global temporal features; the second and third multilayer perceptrons have the same structure, both consisting of linear dimensionality reduction layers and ReLU activation functions. Each layer is followed by a ReLU activation function after linear transformation to maintain nonlinear features; inputting the high-dimensional global temporal features into the second multilayer perceptron compresses the feature dimension from 64 dimensions to 32 dimensions, and inputting the output features of the second multilayer perceptron into the third multilayer perceptron compresses the feature dimension from 32 dimensions to 16 dimensions; the fourth multilayer perceptron contains only a linear output layer, mapping the 16-dimensional features output by the third multilayer perceptron to a 1-dimensional predicted value of lithium battery capacity.

[0037] Preferably, the specific process of S22 is as follows:

[0038] Step 1: Randomly initialize the position and velocity of the particle swarm within a preset parameter space. The position vector of each particle represents a set of parameter combinations including the mode number and penalty factor. At the same time, generate a set of uniformly distributed unit reference vectors in the four-dimensional target space as a guide for the search direction. Establish an external archive to store the set of non-dominated optimal solutions found during the search process.

[0039] Step 2: Substitute the position parameters of each particle into the variational mode decomposition model, decompose the dimensionality-reduced health factor matrix, and calculate four objective function values; map the four objective function values ​​to the target space as the current fitness of the particle;

[0040] Step 3: When updating the individual optimal position of a particle, if the new position is better than the old position in terms of Pareto dominance, then update it directly; if the two do not dominate each other, calculate the distance of the particle to the ideal point in the target space and select the position that is closer as the individual optimal position; when selecting the globally optimal guiding particle, a reference vector association strategy is adopted. Within each region controlled by the reference vector, the non-dominated solution that is closest to the reference vector and has the best convergence is selected as the global leader of the particles in that region.

[0041] Step 4: Update the velocity and position of each particle based on the inertia weight, individual cognitive component, and social group component of the particle swarm optimization algorithm; whereby the inertia weight adopts a nonlinear dynamic adjustment strategy, and performs a polynomial mutation operation on the updated particle position;

[0042] Step 5: At the end of each iteration, update the external archive with the newly generated population; mix the new solutions with the existing solutions in the archive, first removing the dominated inferior solutions; when the number of solutions in the archive exceeds the preset limit, calculate the angular crowding between solutions, and prioritize retaining those solutions that are sparsely distributed in the four-dimensional target space and can maintain the uniformity of population distribution by maximizing the angle, while removing redundant solutions in dense regions.

[0043] Step Six: Determine if the algorithm has reached the maximum number of iterations; if not, return to Step Two and continue the loop; if it has reached the maximum, output a set of Pareto optimal solutions stored in the external archive; using the weighted objective function formula, substitute the four objective function values ​​of each solution in the archive into the calculation, compare the weighted total objective value of all non-dominated solutions, and select the solution with the smallest weighted objective function value as the final optimal number of modes and the best penalty factor.

[0044] Compared with the prior art, the present invention has the following beneficial effects:

[0045] (1) Adaptive Mode Decomposition Strategy Based on Multi-Objective Optimization to Solve the Problem of Key Mode Loss: To address the difficulty in determining the number of modes and penalty factor in traditional variational mode decomposition, a high-dimensional multi-objective particle swarm optimization algorithm guided by reference vector is proposed. This strategy constructs a four-dimensional objective function including weighted correlation coefficient, effective mode ratio, reconstruction error, and average information entropy. The particle swarm algorithm with reference vector is used to search for the Pareto optimal solution set in the parameter space. By adaptively determining the optimal number of modes and penalty factor, the accurate decomposition of multi-dimensional health factors of lithium batteries is achieved. This mechanism effectively solves the problem of key mode loss or over-decomposition caused by fixed parameters in traditional methods, and significantly improves the purity and completeness of signal feature extraction.

[0046] (2) Dual-path spatiotemporal feature coupling and causal enhancement mechanism enhance model generalization ability: In order to fully capture the spatiotemporal dependencies of lithium battery operation data, a dual-path feature extraction architecture including graph attention network and modal projection was constructed. In the first path, seven health factors are constructed as a fully connected graph, and the spatial interaction features between factors are extracted using multi-head graph attention convolutional layers; in the second path, nonlinear projection and dimension matching are performed on the intrinsic modal components. Subsequently, through the causal feature enhancement network, the correlation coefficient between the two features is calculated and the weight vector is generated using the Tanh-activated attention layer. This dual-path fusion design not only explores the global temporal evolution law of the data, but also accurately quantifies the nonlinear coupling strength between different health factors, which greatly enhances the model's generalization ability and noise resistance under complex working conditions;

[0047] (3) Cascaded deep residual prediction architecture and decoupled training strategy significantly improve the accuracy of lithium battery capacity prediction: A composite prediction model consisting of a cascaded two-layer residual network, a long short-term memory network, and a multilayer perceptron decoding was designed, along with a targeted optimization strategy. The cascaded residual blocks are used to solve the gradient vanishing problem in deep networks through a skip connection mechanism, thereby strengthening deep feature extraction. At the same time, an adaptive optimization mechanism that decouples weight decay and gradient update is introduced during the training phase, along with a cosine annealing learning rate scheduling strategy with hot restart. This architecture, combined with the unique training strategy, effectively overcomes the overfitting and local extremum traps of traditional models in long sequence prediction, ensuring that the model can converge to a better solution on the non-convex optimization plane, and significantly reducing the error in capacity prediction. Attached Figure Description

[0048] Figure 1 This is a flowchart illustrating the overall technical route of the present invention.

[0049] Figure 2 This is a structural diagram of a lithium battery operating state feature extraction model.

[0050] Figure 3 This is a structural diagram of the lithium battery capacity causal characteristic enhancement module.

[0051] Figure 4 This is a structural diagram of a lithium battery capacity prediction module.

[0052] Figure 5 This is a comparison chart of the mean absolute errors of all models in the examples.

[0053] Figure 6 This is a comparison chart of the root mean square error of all models in the examples.

[0054] Figure 7 This is a comparison chart of the average absolute percentage error of all models in the examples.

[0055] Figure 8This is a comparison chart of the coefficients of determination for all models in the examples. Detailed Implementation

[0056] This invention proposes a time series prediction method for lithium battery capacity based on dual-path feature fusion and spatiotemporal modeling. The overall technical flowchart is shown below. Figure 1 As shown, the specific steps are as follows:

[0057] S1, Lithium battery health factor data collection; In order to accurately model the lithium battery capacity, this invention collects data on seven health factors that affect the lithium battery capacity, including time-series data of the highest temperature, the lowest temperature, the highest voltage, the lowest voltage, the voltage difference, the current, and the battery pack voltage.

[0058] S2, Construct a lithium battery operating state feature extraction model; this module includes a lithium battery health factor feature decomposition module and a deep residual feature enhancement module; the lithium battery health factor feature decomposition module includes principal component analysis and intrinsic mode function decomposition to obtain the intrinsic mode function components of the lithium battery health factors; the deep residual module includes two feature extraction paths, the first path performs temporal encoding and spatial feature extraction on the lithium battery health factors; the second path performs projection and dimension matching on the intrinsic mode function components of the lithium battery health factors;

[0059] S3, construct a causal feature enhancement module for lithium battery capacity; the causal feature enhancement module for lithium battery capacity includes a graph-structured nonlinear enhancement network, a multilayer perceptron, and a causal feature enhancement network; input the lithium battery health factor graph features output from the first path of the deep residual fusion module into the graph-structured nonlinear enhancement network to obtain enhanced lithium battery health factor graph features; input the intrinsic mode function projection features output from the second path of the deep residual fusion module into the multilayer perceptron to obtain enhanced intrinsic mode function projection features; concatenate the enhanced lithium battery health factor graph features and the enhanced intrinsic mode function projection features and input them into the causal feature enhancement network to obtain the causal feature vector of lithium battery capacity;

[0060] S4, Construct a lithium battery capacity prediction module; the lithium battery capacity prediction module includes a cascaded two-layer residual network, a long short-term memory network, and a lithium battery capacity decoding network; input the lithium battery capacity causal feature vector into the two-layer residual network to obtain enhanced lithium battery capacity causal features; input the enhanced lithium battery capacity causal features into the long short-term memory network to obtain global time-series features; input the global time-series features into the lithium battery capacity decoding network to obtain the predicted lithium battery capacity value.

[0061] The specific implementation process of the present invention will be described in detail below with reference to specific embodiments.

[0062] S1, Lithium Battery Health Factor Data Collection

[0063] S1-1 High and Low Temperature Time Series Data Acquisition: Based on the NTC thermistor array distributed within the lithium battery module, the sampling algorithm traverses all probe data at each sampling moment to obtain the highest and lowest temperature values ​​at each moment in the time window, thereby forming the highest temperature time series data and the lowest temperature time series data.

[0064] S1-2 Acquisition of time-series data for highest voltage, lowest voltage, and voltage difference: The voltage of all individual cells is acquired in parallel using an analog front-end. The sampling algorithm selects the maximum and minimum values ​​of the individual cell voltage at each moment in the time window and calculates the difference to obtain the instantaneous voltage difference, which in turn forms the time-series data for highest voltage, lowest voltage, and voltage difference.

[0065] S1-3 Current and battery pack voltage timing data acquisition: Within the time window, the current time series and battery pack voltage time series are acquired by the Hall sensor and total voltage detection circuit on the main circuit of the lithium battery module.

[0066] S2. Construct a lithium battery operating state feature extraction model

[0067] Model structure as follows Figure 2 As shown, it includes a lithium battery health factor feature decomposition module and a deep residual feature enhancement module; the lithium battery health factor feature decomposition module includes principal component analysis and intrinsic mode function decomposition to obtain the intrinsic mode function components of the lithium battery health factor; the deep residual module includes two feature extraction paths, the first path performs temporal encoding and spatial feature extraction on the lithium battery health factor; the second path projects and matches the eigenmode function components of the lithium battery health factor.

[0068] S2-1 Lithium Battery Health Factor Characteristic Decomposition Module:

[0069] 1. First, principal component analysis was performed on the lithium battery health factor data obtained from S1, as follows:

[0070] 1) Combine the highest temperature time-series data, lowest temperature time-series data, highest voltage time-series data, lowest voltage time-series data, voltage difference time-series data, current time-series data, and battery pack voltage time-series data obtained from S1 to form a health factor matrix. ,in For the sample size, For the number of factors; Each sample in the dataset is standardized:

[0071] ;

[0072] in Indicates the first The first sample The standardized values ​​of each health factor Indicates the first The first sample The original values ​​of each health factor Indicates the first The mean of each health factor Indicates the first The standard deviation of each health factor is used to obtain the standardized health factor matrix. .

[0073] 2) Calculate the covariance matrix And perform eigenvalue decomposition:

[0074] ;

[0075] in This represents the covariance matrix of the standardized health factor matrix. This represents the standardized health factor matrix. Indicates the first 1 eigenvector Indicates the first Each feature value.

[0076] 3) Sort the feature values ​​in descending order and select the top [values]. The eigenvectors form the projection matrix; multiplying the health factor matrix with the projection matrix yields the dimensionality-reduced health factor matrix.

[0077] 2. Secondly, an intrinsic mode function decomposition model based on multi-objective optimization is constructed to perform variational mode decomposition on the dimensionality-reduced health factor matrix, as follows:

[0078] 1) Construct a multi-objective optimization problem for the intrinsic mode function decomposition of lithium battery health factors, as detailed below:

[0079] ;

[0080] ;

[0081] in This represents the weighted objective function. , , , They represent the objective functions respectively. Weighting coefficients; first objective function Used to evaluate the correlation between intrinsic mode function components and the original signal. The weighted correlation coefficient is calculated using the following formula:

[0082] ;

[0083] in This represents the set of effective intrinsic mode functions with an absolute correlation coefficient greater than 0.3. Indicates the first The correlation coefficients between each intrinsic mode function component and the original signal This represents the number of effective intrinsic mode functions.

[0084] Second objective function Used to evaluate the proportion of effective intrinsic mode functions. The effective intrinsic mode function proportion represents the percentage of intrinsic mode functions whose correlation coefficient exceeds a threshold. The calculation formula is as follows:

[0085] ;

[0086] in This represents the total number of modes in the VMD decomposition.

[0087] Third objective function Used to measure the accuracy of signal reconstruction after variational mode decomposition, reconstruction error The calculation formula is:

[0088] ;

[0089] in Indicates the original fusion health factor sequence. Indicates the first Each intrinsic mode function component Denotes the squared L2 norm of a vector;

[0090] Fourth objective function Reflects the average complexity of the intrinsic mode function components. The average information entropy is calculated using the following formula:

[0091] ;

[0092] in Indicates the first The information entropy of each intrinsic mode function component reflects the complexity of that component;

[0093] The decision variable is the number of modes in variational mode decomposition. and penalty factor .

[0094] 2) A high-dimensional multi-objective particle swarm optimization algorithm based on reference vector guidance and inflection point optimization strategy is designed to solve the multi-objective optimization problem of intrinsic mode function decomposition of lithium battery health factors. This algorithm introduces a set of predefined reference vectors to guide the particle swarm to a uniform distribution in four-dimensional space, and combines angle penalty distance for environment selection, ultimately outputting the optimal number of modes. and optimal penalty factor The variational mode decomposition method with optimal mode number and optimal penalty factor is used to perform variational mode decomposition on the lithium battery health factor to obtain the intrinsic mode function components of the lithium battery health factor.

[0095] The specific steps include:

[0096] Step 1: Algorithm Initialization and Reference Vector Construction; First, the positions and velocities of the particle swarm are randomly initialized within a predefined parameter space, where the position vector of each particle represents a set of parameter combinations including the mode number and penalty factor; Simultaneously, to address the high-dimensional challenge posed by the four objective functions, a set of uniformly distributed unit reference vectors is generated in the four-dimensional objective space. These vectors will serve as guides for the search direction, helping the population cover the entire Pareto front; In addition, an external archive is established to store the set of non-dominated optimal solutions discovered during the search process;

[0097] Step 2: Fitness Assessment and Target Space Mapping; Substitute the position parameters of each particle into the variational mode decomposition model to decompose the dimensionality-reduced health factor matrix, and calculate the aforementioned four objective function values, including weighted correlation coefficient, effective mode ratio, reconstruction error, and average information entropy; Then map these four objective function values ​​to the target space as the particle's current fitness.

[0098] Step 3: Individual and Global Optimal Updates Based on Reference Vectors; When updating the individual optimal position of a particle, if the new position is better than the old position in terms of Pareto dominance, it is updated directly; if the two are not mutually dominant, the distance of the particle to the ideal point in the target space is calculated, and the position with the closer distance, i.e., better convergence, is selected as the individual optimal position; When selecting the globally optimal guiding particle, a reference vector association strategy is adopted, that is, the angle between each particle and all reference vectors is calculated, and the particle is associated with the region to which the reference vector with the smallest angle belongs; Within each region controlled by a reference vector, the non-dominated solution that is closest to the reference vector and has the best convergence is selected as the global leader of the particles in that region; This strategy ensures that the particle swarm not only converges to the true frontier, but also maintains diversity in the trade-offs between the four objectives, avoiding premature convergence of the algorithm;

[0099] Step 4: Position and velocity update and mutation operation; The velocity and position of each particle are updated according to the inertia weight, individual cognitive component and social group component of the traditional particle swarm optimization algorithm; Among them, the inertia weight adopts a non-linear dynamic adjustment strategy, which maintains a large value in the early stage of iteration to enhance the global exploration capability, and decreases it in the later stage to improve the local exploration accuracy; In order to prevent getting trapped in local extrema, a polynomial mutation operation is performed on the updated particle position, that is, the mode number or penalty factor is slightly perturbed with a certain probability, thereby increasing the diversity of the population;

[0100] Step 5: External archive maintenance based on angle penalty distance; at the end of each iteration, update the external archive using the newly generated population; mix the new solutions with the existing solutions in the archive, first removing the dominated inferior solutions; when the number of solutions in the archive exceeds the preset limit, execute the environment selection mechanism based on angle penalty distance, that is, calculate the angle crowding between solutions, prioritize retaining those solutions that are sparsely distributed in the four-dimensional target space and can maintain the uniformity of population distribution by maximizing the angle, and remove redundant solutions in dense areas;

[0101] Step Six: Select the optimal parameters based on weighted decision-making; determine if the algorithm has reached the maximum number of iterations; if not, return to Step Two and continue the loop; if it has reached the maximum, output a set of Pareto optimal solutions stored in the external archive; finally, using the weighted objective function formula proposed in this invention, substitute the four objective function values ​​of each solution in the archive into the calculation, compare the weighted total objective value of all non-dominated solutions, and select the mode number and penalty factor corresponding to the solution with the smallest weighted objective function value as the final optimal mode number and optimal penalty factor.

[0102] S2-2 Constructing the deep residual module:

[0103] The deep residual module contains two feature extraction paths. The first path performs temporal encoding and spatial feature extraction on lithium battery health factors. The second path projects and performs dimension matching on the intrinsic mode function components of lithium battery health factors obtained by S2-1.

[0104] 1. The first path of the deep residual module is used to perform temporal encoding and spatial feature extraction of lithium battery health factors, as detailed below:

[0105] 1) Construct a time series encoder to encode lithium battery health factors in a time series; the encoder includes an input layer, a regularization layer, a ReLU activation function layer, and a random deactivation layer; input the lithium battery health factors into the encoder to reduce the dimensionality and obtain the time series encoded features of the lithium battery health factors; the regularization layer and the random deactivation layer are used to enhance the generalization ability of the model;

[0106] 2) Construct a spatial feature extraction module for lithium battery health factors;

[0107] First, construct a fully connected graph of lithium battery health factors. ,in , Health factors are used as nodes. Health factor data serves as node features, and the interactions between health factors serve as edges connecting two health factors. The side lengths represent the magnitude of the influence between health factors, and all side lengths are set to 1; this graph structure is used to model the spatial relationships between health factors.

[0108] Secondly, a spatial feature extraction network for lithium battery health factors is constructed to calculate the weights of each edge in the fully connected graph of lithium battery health factors and extract the spatial features of health factors. This network includes a multi-head graph attention convolutional layer, a ReLU activation function layer, a random deactivation layer, a single-head graph attention convolutional layer, and a graph aggregation layer. First, the features of each node in the fully connected graph of lithium battery health factors are input into the multi-head graph attention convolutional layer to obtain the multi-head spatial attention features of each node in the fully connected graph. The features of all nodes constitute the multi-head spatial attention features of battery health factors. Second, the multi-head spatial attention features of lithium battery health factors output by the multi-head graph attention convolutional layer are input into the ReLU activation function layer for feature activation, enhancing the expression of features related to lithium battery life and suppressing irrelevant features. Third, the features output by the ReLU activation function layer are input into the random deactivation layer to suppress the expression of some neurons in the model, reduce the risk of overfitting, and obtain the high-level spatial features of lithium battery health factors.

[0109] Secondly, the high-level features of the lithium battery health factor space are input into a single-headed graph attention convolutional layer to calculate the attention coefficient and weight the channel feature values ​​for each feature channel, thereby obtaining the deep features of the lithium battery health factor space. The deep features of the lithium battery health factor space are then input into a regularization layer to perform layer normalization on the feature values, preventing gradient direction loss due to excessive differences in the order of magnitude of the lithium battery health factor feature values. The normalized deep features of the lithium battery health factor space are then input into a random deactivation layer and then into a graph aggregation layer to calculate the mean of each feature channel and update the features of each node in the fully connected graph of the lithium battery health factor, thereby obtaining the graph features of the lithium battery health factor.

[0110] 2. The second path of the deep residual module is used to perform temporal feature projection and dimensionality matching on the intrinsic mode function components of the lithium battery health factor, as detailed below:

[0111] First, the intrinsic mode function (IMF) components obtained in S2-1 are linearly projected into the input layer of the temporal mapping network to map the feature dimension to 64 dimensions, and then input into the normalization layer for standardization. The normalized features are then input into the ReLU activation function layer for nonlinear activation to enhance the nonlinear expressive power of the IMF features. Next, the features are input into the random deactivation layer to randomly suppress the expression of some neurons, enhancing the model's generalization ability and reducing the risk of overfitting, thus obtaining intermediate features of the IMF components. Second, the intermediate features of the IMF components are input into the output layer of the multilayer perceptron for secondary dimensional projection, further mapping the feature dimension to 32 dimensions to match the feature dimension of the lithium battery health factor map obtained by the first path of the deep residual module. The mapped features are then input into the normalization layer to perform layer normalization on the feature values ​​to prevent excessive differences in the order of magnitude of the feature values ​​from causing gradient anomalies, thus obtaining the projected features of the IMF, providing feature input for subsequent multi-scale attention fusion.

[0112] S3. Construct a causal feature enhancement module for lithium battery capacity.

[0113] The structure of the lithium battery capacity causal characteristic enhancement module is as follows: Figure 3 As shown, the system includes a graph-structured nonlinear enhancement network, a multilayer perceptron, and a causal feature enhancement network. The lithium battery health factor graph features output from the first path of the deep residual fusion module are input into the graph-structured nonlinear enhancement network to obtain enhanced lithium battery health factor graph features. The intrinsic mode function projection features output from the second path of the deep residual fusion module are input into the multilayer perceptron to obtain enhanced intrinsic mode function projection features. The enhanced lithium battery health factor graph features and the enhanced intrinsic mode function projection features are concatenated and simultaneously input into the causal feature enhancement network to obtain the lithium battery capacity causal feature vector.

[0114] First, the lithium battery health factor graph features output from the first path of the deep residual fusion module are input into a graph-structured nonlinear enhancement network. Simultaneously, the intrinsic mode function projection features output from the second path are input into a multilayer perceptron for nonlinear mapping and feature enhancement. This yields enhanced lithium battery health factor graph features and intrinsic mode function features, preserving the expressive power of the features before fusion. The graph-structured nonlinear enhancement network consists of stacked graph convolutional layers and ReLU activation function layers, with each graph convolutional layer followed by a ReLU activation function layer. The multilayer perceptron consists of single fully connected layers.

[0115] Secondly, a causal feature enhancement network is constructed. First, the enhanced lithium battery health factor map features and intrinsic mode function (IMF) projection features are concatenated along the feature dimension to form a joint feature vector. Next, the joint feature vector is input into an attention layer containing a Tanh activation function to calculate the correlation coefficient. This correlation coefficient quantifies the influence between each channel of the lithium battery health factor map features and each channel of the IMF projection features, i.e., the causal relationship between the two types of features. Then, the coefficients are normalized to a probability distribution using the Softmax function, thereby generating an attention weight vector containing the weights of both feature paths. Finally, a weighted feature fusion operation is performed. The calculated attention weight vector is used to perform element-wise weighted multiplication on the enhanced lithium battery health factor map features and the IMF projection features, respectively, to achieve adaptive feature selection, obtaining weighted lithium battery health factor map features and weighted IMF projection features. The weighted features are then concatenated again to form a causal feature vector for lithium battery capacity, which serves as the input to the subsequent lithium battery capacity prediction module. This module maintains the batch and sequence dimensions unchanged throughout the entire process, achieving adaptive fusion of feature dimensions only through the attention mechanism.

[0116] S4. Construct a lithium battery capacity prediction module

[0117] Lithium battery capacity prediction module, such as Figure 4 As shown, it includes a cascaded two-layer residual network, a long short-term memory network, and a lithium battery capacity decoding network; the lithium battery capacity causal feature vector is input into the two-layer residual network to obtain enhanced lithium battery capacity causal features; the enhanced lithium battery capacity causal features are input into the long short-term memory network to obtain global temporal features; and the global temporal features are input into the lithium battery capacity decoding network to obtain the predicted lithium battery capacity value.

[0118] S4-1 constructs a cascaded two-layer residual network structure: This network contains two residual blocks, which perform deep feature extraction on the causal feature vector of lithium battery capacity obtained by S4; the calculation process of each residual block is as follows: First, the input features are passed through a linear transformation layer, a normalization layer, and a ReLU activation function layer in sequence for feature mapping and nonlinear activation; then, through a skip connection mechanism, the original input features of the residual block are directly superimposed on the transformed main path features. If the input and output dimensions do not match, the input features are linearly projected to align the dimensions, so as to prevent the gradient vanishing problem in the process of network deepening and to preserve the original information;

[0119] The output of the first residual block is used as the input of the second residual block, and the above operation process is repeated to finally obtain the causal features of the lithium battery capacity after deep residual enhancement, which are then used as the input of the subsequent long short-term memory network.

[0120] S4-2 Constructs a unidirectional, single-layer long short-term memory network: Using the enhanced lithium battery capacity causal features output by the deep residual fusion module as input, the gating mechanism in the long short-term memory network is used to recursively update the cell state and hidden state at each time step, capturing the long-term temporal dependencies in the lithium battery capacity decay process, and extracting the hidden state of the last time step of the sequence as the global temporal feature.

[0121] S4-3 Constructing a Lithium-ion Battery Capacity Decoding Network: The network consists of four layers of a multilayer perceptron, mapping the global temporal features output by the Long Short-Term Memory (LSTM) network to predicted lithium-ion battery capacity values. Specifically, the first perceptron layer includes a linear input layer, a normalization layer, a ReLU activation function layer, and a random deactivation layer. The linear input layer performs a linear mapping on the global temporal features, increasing the feature dimension and obtaining richer high-dimensional features of lithium-ion battery degradation. These high-dimensional features are then sequentially input into the normalization layer, the ReLU activation function layer, and the random deactivation layer, where feature value normalization, nonlinear activation, and random neuron deactivation operations are performed, respectively. The high-dimensional global temporal features are obtained. The second and third multilayer perceptrons have the same structure, consisting of linear dimensionality reduction layers and ReLU activation functions. Each layer is followed by a ReLU activation function after linear transformation to maintain nonlinear features. The high-dimensional global temporal features are input into the second multilayer perceptron to compress the feature dimension from 64 dimensions to 32 dimensions. The output features of the second multilayer perceptron are input into the third multilayer perceptron to compress the feature dimension from 32 dimensions to 16 dimensions. The fourth multilayer perceptron contains only a linear output layer and maps the 16-dimensional features output by the third multilayer perceptron to a 1-dimensional lithium battery capacity prediction value.

[0122] S5, Model Training

[0123] The training strategy designed in this invention aims to implement a targeted training strategy for the complex network structure constructed in this invention, so as to minimize the error between the predicted value and the true value of lithium battery capacity. The specific steps are as follows:

[0124] S5-1 is configured to adapt to a decoupled weight decay adaptive optimization mechanism for multi-source feature fusion networks: For the heterogeneous network structures included in the lithium battery operating state feature extraction model, the lithium battery capacity causal feature enhancement module, and the lithium battery capacity prediction module in this invention, an adaptive moment estimation optimizer with decoupled weight decay and gradient update is designed for parameter updates. During training, this optimizer first calculates the first-order and second-order gradient moment estimates of the network parameters based on the loss function of lithium battery capacity prediction, to dynamically adjust the learning step size of the deep residual module, the graph structure nonlinear enhancement network, the causal feature enhancement network, the cascaded two-layer residual network, the long short-term memory network, and the lithium battery capacity decoding network. Simultaneously, to overcome the overfitting problem easily caused by the high feature dimension during dual-path feature fusion, this mechanism independently applies decay penalties to the network weights, directly constraining the temporal encoding and spatial feature extraction paths in the deep residual module and the multilayer perceptron parameter amplitudes in the lithium battery capacity decoding network, without mixing them into the gradient calculation.

[0125] S5-2 designs a cosine annealing hot-restart scheduling strategy for the non-convex optimization plane of capacity prediction: Considering the nonlinearity and multi-factor coupling characteristics of the lithium battery capacity decay process, its loss function surface is extremely uneven. Therefore, a cosine annealing learning rate scheduling strategy with a hot-restart mechanism is introduced. During the training cycle, the learning rate gradually decays from its maximum value according to the cosine function curve, so that the model can quickly traverse the parameter space of the lithium battery operating state feature extraction model and the lithium battery capacity causal feature enhancement module in the early stage. When a cycle ends, the learning rate is instantly reset to a high value. Through repeated exploration with large step size and convergence with small step size, it is ensured that the cascaded two-layer residual network, long short-term memory network and lithium battery capacity decoding network can converge to a better parameter combination, thereby improving the fitting accuracy of the capacity evolution trend of the lithium battery throughout its entire life cycle.

[0126] S5-3 establishes an early termination and model solidification process based on capacity prediction error monitoring: Monitoring indicators and termination conditions are set for the training process; during iteration, the mean square error of lithium battery capacity prediction on the validation set is calculated in real time, and model performance is monitored. If the capacity prediction error on the validation set fails to decrease further over several consecutive cycles, the entire training process for the lithium battery operating state feature extraction model, the lithium battery capacity causal feature enhancement module, and the lithium battery capacity prediction module will be terminated, and the network parameters at the moment with the lowest validation error will be solidified as the final model. This process ensures that the model fully learns the lithium battery capacity degradation law while effectively avoiding the memorization of training set noise caused by excessive iteration.

[0127] Simulation experiments and results analysis:

[0128] This embodiment trains the model on 16 randomly numbered lithium battery samples and tests it on the remaining 4 batteries, comparing it with existing mainstream machine learning and deep learning models. The comparison models include: Transformer, Informer, optimized TCN, Bidirectional LSTM (BiLSTM), VMD-LSTM, GAT-LSTM, LSTM, and MTGNN. Evaluation metrics include MAE, RMSE, MAPE, and R².

[0129] ;

[0130] ;

[0131] ;

[0132] ;

[0133] in Mean absolute error measures the average absolute difference between predicted and actual values. The root mean square error (RMSE) measures the standard deviation between the predicted and actual values. It represents the mean absolute percentage error, measuring the average of the relative errors in prediction. The coefficient of determination measures how well the model explains the variation in the target variable. Indicates the first The actual battery capacity value of each sample Indicates the first The predicted battery capacity value for each sample. This represents the average actual battery capacity of all samples. This represents the total number of samples.

[0134] Experimental results are as follows Figures 5 to 8 As shown, the following conclusions can be drawn:

[0135] 1. Optimal Overall Prediction Accuracy: The method proposed in this invention, based on dual-path feature fusion and spatiotemporal modeling, achieves the best performance across all evaluation metrics. Its mean absolute error (MAE) is only 1.626, root mean square error (RMSE) is 2.043, and mean absolute percentage error (MAPE) is as low as 1.34%. Compared to the second-best performing Informer model, the method of this invention reduces MAE by approximately 5.2% and RMSE by approximately 13.1%, indicating that this invention has higher accuracy in capturing the capacity decay trend of lithium batteries.

[0136] 2. Significantly Enhanced Fitting Ability: The coefficient of determination (R²) of this invention reached 0.829, which is 14.8% and 13.7% higher than that of traditional LSTM and BiLSTM, respectively. This indicates that by introducing a dual-path feature fusion mechanism, combining the spatial features of the graph structure with the frequency domain features of the intrinsic mode functions, the problem of insufficient fitting of a single model when dealing with nonlinear capacity decay curves can be effectively solved.

[0137] 3. Outperforms mainstream deep learning models: Although models based on self-attention mechanisms such as Transformer and Informer perform well in time series prediction (R² 0.790 and 0.788, respectively), the method of this invention still maintains a significant leading advantage. This is mainly due to the deep residual module and causal feature enhancement network designed in this invention, which not only captures long-term time series dependencies but also mines the complex causal relationships between health factors, thus showing stronger robustness in the face of fluctuations and noise in lithium battery capacity data.

[0138] 4. Advantages compared to combined models: Compared with the two simple combined models, VMD+LSTM and GAT+LSTM, the prediction error of the method in this invention is significantly reduced. This verifies the effectiveness of the VMD decomposition strategy based on multi-objective optimization and the graph structure nonlinear enhancement network adopted in this invention, and proves that a well-designed feature fusion architecture can better mine the deep value of data than simple model stacking.

[0139] In summary, the method of this invention significantly outperforms existing mainstream time-series prediction models in terms of both prediction accuracy and model interpretability in the task of predicting lithium battery capacity, verifying the effectiveness and advancement of this method in the field of battery health status assessment.

[0140] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

[0141] While the specific embodiments of the present invention have been described above, they are not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A lithium battery capacity prediction method based on two-path feature fusion and spatio-temporal modeling, characterized in that, The method comprises the following steps: S1, collecting time window lithium battery health factor time series data; S2, inputting the collected data into a lithium battery operating state feature extraction model, the model comprising a lithium battery health factor feature decomposition module and a deep residual module; the lithium battery health factor feature decomposition module comprising principal component analysis and intrinsic mode function decomposition, and obtaining lithium battery health factor intrinsic mode function components; the deep residual module comprising two feature extraction paths, the first path performing time series coding and spatial feature extraction on the lithium battery health factor time series data to obtain lithium battery health factor graph features; the second path projecting and dimension matching the lithium battery health factor intrinsic mode function components to obtain intrinsic mode function projection features; The lithium battery health factor feature decomposition module specifically comprises: firstly, performing principal component analysis on the obtained lithium battery health factor time series data to obtain a reduced dimension health factor matrix; then, constructing a intrinsic mode function decomposition model based on multi-objective optimization to perform variational mode decomposition on the reduced dimension health factor matrix, comprising: S21, constructing a lithium battery health factor EMD multi-objective optimization problem; the lithium battery health factor EMD multi-objective optimization problem includes a multi-objective weighted objective function and a decision variable; the multi-objective weighted objective function is calculated by weighted summation of four objective functions; the first objective function is used to evaluate the correlation between the EMD component and the original signal, which is quantified by a weighted correlation coefficient; the second objective function is used to evaluate the proportion of effective EMD, which is quantified by the proportion of effective EMD; the third objective function is used to measure the accuracy of signal reconstruction after variational mode decomposition, which is quantified by reconstruction error; the fourth objective function reflects the average complexity of the EMD component, which is quantified by the average information entropy; the decision variable is the mode number in the variational mode decomposition and a penalty factor ; S22, based on the reference vector guide and the inflection point optimization strategy of high-dimensional multi-objective particle swarm optimization algorithm to solve lithium battery health factor intrinsic mode function decomposition multi-objective optimization problem; A set of predefined reference vectors are introduced into the traditional particle swarm optimization algorithm to guide the uniform distribution of particle swarm in four-dimensional space, and the angle penalty distance is combined with the environment selection, and finally the optimal modal number is output and the optimal penalty factor ; The optimal modal number and the optimal penalty factor are used to configure the variational modal decomposition method to decompose the lithium battery health factor, and the lithium battery health factor intrinsic mode function component is obtained S3, enhancing the features obtained in S2 based on a lithium battery capacity causal feature enhancement module; inputting the lithium battery health factor graph features into a graph structure nonlinear enhancement network to obtain enhanced lithium battery health factor graph features; inputting the intrinsic mode function projection features into a multi-layer perception machine to obtain enhanced intrinsic mode function projection features; inputting the two enhanced features into a causal feature enhancement network to obtain a lithium battery capacity causal feature vector; S4, inputting the lithium battery capacity causal feature vector into a lithium battery capacity prediction module to obtain a lithium battery capacity prediction value.

2. The lithium battery capacity prediction method based on dual-path feature fusion and spatio-temporal modeling of claim 1, wherein: The lithium battery health factor time series data comprises: the highest temperature and the lowest temperature of all probes at each time point in the time window, forming highest temperature time series data and lowest temperature time series data; the maximum and minimum values of the single cell voltage at each time point in the time window, and the instantaneous voltage difference is obtained by subtracting the minimum value from the maximum value, and then the highest voltage time series data, the lowest voltage time series data and the voltage difference time series data are formed; the current time series and the battery pack voltage time series on the main loop of the lithium battery module in the time window.

3. The lithium battery capacity prediction method based on dual-path feature fusion and spatio-temporal modeling of claim 1, wherein: The first path of the deep residual module specifically comprises the following data processing process: firstly, constructing a time series encoder to perform time series coding on the lithium battery health factor; the encoder comprises an input layer, a regularization layer, a ReLU activation function layer and a random inactivation layer; inputting the lithium battery health factor into the encoder to obtain lithium battery health factor time series coding features; then, constructing a lithium battery health factor spatial feature extraction module: Building a full connection graph of lithium battery health factors Wherein , ; wherein the health factors are nodes , the health factor data is the node feature, and the interaction between the health factors is the edge connecting two health factors , the edge length represents the influence size between the health factors, and all the edge lengths are set to 1 constructing a lithium battery health factor spatial feature extraction network to calculate the weight of each edge in the lithium battery health factor full connection graph and extract health factor spatial features; the network comprises a multi-head graph attention convolution layer, a ReLU activation function layer, a random inactivation layer, a single-head graph attention convolution layer and a graph aggregation layer; firstly, inputting the features of each node in the lithium battery health factor full connection graph into the multi-head graph attention convolution layer to obtain the multi-head spatial attention features of each node in the full connection graph, and the features of all nodes form lithium battery health factor multi-head spatial attention features; Secondly, the lithium battery health factor multi-head spatial attention feature is input into a ReLU activation function layer for feature activation; thirdly, the features output by the ReLU activation function layer are input into a random inactivation layer to suppress the expression of part of the neurons of the model, so as to obtain lithium battery health factor spatial high-level features; The lithium battery health factor spatial high-level features are input into a single-head graph attention convolution layer to calculate attention coefficients and weight channel feature values for each feature channel, so as to obtain lithium battery health factor spatial depth features; The lithium battery health factor spatial depth features are input into a regularization layer for layer normalization operation on the feature values; the normalized lithium battery health factor spatial depth features are input into a random inactivation layer and then input into a graph aggregation layer, the mean value of each feature channel is calculated, and the features of each node in the lithium battery health factor fully connected graph are updated, so as to obtain lithium battery health factor graph features.

4. The lithium battery capacity prediction method based on dual-path feature fusion and spatio-temporal modeling of claim 1, wherein: The specific data processing process of the second path of the deep residual module is as follows: The obtained intrinsic modal function component is input into the input layer of the time sequence mapping network for linear projection, the feature dimension is mapped to 64 dimensions, and the normalized layer is input for standardization processing; The normalized features are input into a ReLU activation function layer for nonlinear activation, and then input into a random inactivation layer to randomly suppress the expression of part of the neurons, so as to obtain intrinsic modal function component intermediate features; secondly, the intrinsic modal function component intermediate features are input into the output layer of a multi-layer perceptron for secondary dimension projection, the feature dimension is mapped to 32 dimensions to match the feature dimension of the lithium battery health factor graph features obtained by the first path; the mapped features are input into a normalization layer for layer normalization operation on the feature values, so as to obtain intrinsic modal function projection features.

5. The lithium battery capacity prediction method based on dual-path feature fusion and spatio-temporal modeling of claim 1, wherein: The lithium battery capacity causal feature enhancement module specifically comprises: Firstly, the lithium battery health factor graph features are input into a graph structure nonlinear enhancement network, and the intrinsic modal function projection features are input into a multi-layer perceptron for nonlinear mapping and feature enhancement, so as to obtain enhanced lithium battery health factor graph features and intrinsic modal function features; The graph structure nonlinear enhancement network is stacked by a graph convolution layer and a ReLU activation function layer, and each graph convolution layer is connected with a ReLU activation function layer; the multi-layer perceptron is composed of a single full connection layer; Secondly, the enhanced lithium battery health factor graph features and the intrinsic modal function projection features are spliced in the feature dimension to form a joint feature vector; the joint feature vector is input into an attention layer containing a Tanh activation function to calculate a correlation coefficient; the correlation coefficient is normalized into a probability distribution by a Softmax function to generate an attention weight vector containing two feature weight channels; finally, a weighted feature fusion operation is performed, and the attention weight vector is used to perform element-wise weighted multiplication on the enhanced lithium battery health factor graph features and the intrinsic modal function projection features, so as to obtain weighted lithium battery health factor graph features and weighted intrinsic modal function projection features; The two weighted features are spliced again to form a lithium battery capacity causal feature vector.

6. The lithium battery capacity prediction method based on dual-path feature fusion and spatio-temporal modeling of claim 1, wherein: The lithium battery capacity prediction module comprises a cascaded double-layer residual network, a long short-term memory network and a lithium battery capacity decoding network; the lithium battery capacity causal feature vector is input into the double-layer residual network to obtain enhanced lithium battery capacity causal features; the enhanced lithium battery capacity causal features are input into the long short-term memory network to obtain global timing features; The global timing features are input into the lithium battery capacity decoding network to obtain a lithium battery capacity prediction value.

7. The lithium battery capacity prediction method based on dual-path feature fusion and spatio-temporal modeling of claim 6, wherein: The cascaded double-layer residual network comprises two residual blocks for deep feature extraction of the obtained lithium battery capacity causal feature vector; the calculation process of each residual block is as follows: firstly, the input features are sequentially subjected to linear transformation, normalization and ReLU activation function layers for feature mapping and nonlinear activation; then, the original input features of the residual block are directly superimposed on the transformed main path features through a skip connection mechanism, and if the input and output dimensions do not match, the input features are linearly projected to align the dimensions; the output of the first residual block is taken as the input of the second residual block, and the above calculation process is repeated, so that the enhanced lithium battery capacity causal features are finally obtained through deep residual enhancement.

8. The lithium battery capacity prediction method based on dual-path feature fusion and spatio-temporal modeling of claim 6, wherein: The lithium battery capacity decoding network is composed of four multilayer perceptron layers, which map the global timing features output by the long short-term memory network into a lithium battery capacity prediction value; specifically, the first multilayer perceptron comprises a linear input layer, a normalization layer, a ReLU activation function layer and a random inactivation layer, the linear input layer performs linear mapping on the global timing features to obtain high-dimensional features with richer lithium battery degradation characteristics, and the high-dimensional features are sequentially input into the normalization layer, the ReLU activation function layer and the random inactivation layer for feature value normalization, nonlinear activation and neuron random inactivation operations, respectively, to obtain high-dimensional global timing features; The second multilayer perceptron and the third multilayer perceptron have the same structure and are both composed of a linear dimension reduction layer and a ReLU activation function, and the linear transformation of each layer is followed by a ReLU activation function to maintain nonlinear features; The high-dimensional global timing features are input into the second multilayer perceptron to compress the feature dimension from 64 dimensions to 32 dimensions, and the output features of the second multilayer perceptron are input into the third multilayer perceptron to compress the feature dimension from 32 dimensions to 16 dimensions; the fourth multilayer perceptron only comprises a linear output layer, which maps the 16-dimensional features output by the third multilayer perceptron into a 1-dimensional lithium battery capacity prediction value.

9. The lithium battery capacity prediction method based on dual-path feature fusion and spatio-temporal modeling of claim 1, wherein: The specific process of S22 is as follows: Step one: randomly initialize the position and speed of the particle swarm in the preset parameter space, and the position vector of each particle represents a group of parameter combinations containing the modal number and the penalty factor; at the same time, a group of uniformly distributed unit reference vectors are generated in the four-dimensional target space as the guide of the search direction; an external archive is established to store the non-dominated optimal solution set found in the search process; Step two: substitute the position parameters of each particle into the variational modal decomposition model, decompose the reduced health factor matrix, and calculate the values of the four objective functions; map the values of the four objective functions to the target space as the current fitness of the particles; Step three: when updating the individual optimal position of the particle, if the new position is superior to the old position in the Pareto dominance relationship, the new position is directly updated; if the two positions are not dominated by each other, the distance of the particle to the ideal point in the target space is calculated, and the position with shorter distance is selected as the individual optimal position; In the selection of the global optimal guiding particle, the reference vector association strategy is adopted, and in each reference vector controlled area, the non-dominated solution with the shortest distance to the reference vector and the best convergence is selected as the global leader of the particle in the area; Step four: update the speed and position of each particle according to the inertia weight, individual cognitive component and social group component of the particle swarm algorithm; wherein the inertia weight adopts a nonlinear dynamic adjustment strategy, and a polynomial mutation operation is performed on the updated particle position; Step five: at the end of each iteration, the external archive is updated by using the newly generated population; the new solution is mixed with the existing solutions in the archive, and the dominated inferior solutions are first removed; When the number of solutions in the archive exceeds the preset upper limit, the angle crowding degree between solutions is calculated, and the solutions that are sparsely distributed in the four-dimensional target space and can maintain the uniformity of the population distribution by maximizing the angle are preferentially retained, and the redundant solutions in the dense area are removed; Step six: judge whether the algorithm reaches the maximum iteration number; if not, return to step two and continue the loop; if yes, output a set of Pareto optimal solution set stored in the external archive; use the weighted objective function formula to calculate the four objective function values of each solution in the archive, compare the weighted total objective values of all non-dominated solutions, select the modal number and the penalty factor corresponding to the solution with the minimum weighted objective function value as the final optimal modal number and the best penalty factor.

Images