Lithium-ion battery state of health prediction method and device based on st-gnn

Abstract

The application belongs to the technical field of battery health management and artificial intelligence, and relates to a lithium ion battery health state prediction method based on a space-time graph neural network, which comprises the following steps: adopting a SOH prediction model with training convergence to collect electric characteristic data in the charging and discharging process of a lithium ion battery, optimizing a graph network structure and a time sequence feature extraction network based on lithium ion battery charging and discharging physical signal characteristics, accurately capturing space-time correlation characteristics in the battery aging process, and significantly improving prediction accuracy, robustness and generalization, so that the method is suitable for battery SOH prediction in multiple scenes such as power batteries and energy storage power stations.

Description

Technical Field

[0001] This invention belongs to the field of battery health management and artificial intelligence technology, specifically a method and device for predicting the health status of lithium-ion batteries based on spatiotemporal graph neural network ST-GNN, which is applicable to the prediction of the health status of lithium-ion batteries in various scenarios such as power batteries for new energy vehicles, batteries for energy storage power stations, and batteries for consumer electronics. Background Technology

[0002] Driven by the national "dual-carbon" strategy, lithium-ion batteries are widely used in new energy vehicles, energy storage power stations, and consumer electronics, becoming a core infrastructure for the upgrading of the energy industry. However, during long-term cyclic use, batteries are subject to irreversible degradation due to factors such as materials and operating conditions, including cathode material degradation, electrolyte decomposition, and SEI film thickening. This not only shortens lifespan and increases costs but also easily leads to safety hazards such as thermal runaway, hindering the development of the industry.

[0003] Accurate state of health (SOH) estimation is a core technology for safe battery operation and intelligent management. Many national policies have clearly required the establishment of an online monitoring system for battery health, which places high demands on the accuracy, real-time performance, and robustness of SOH prediction algorithms.

[0004] Existing SOH prediction methods are categorized into direct measurement methods, model-driven methods, data-driven methods, and fusion methods, with data-driven methods becoming the mainstream due to their strong feature learning capabilities and adaptability to different operating conditions. However, existing data-driven methods still have significant shortcomings: First, the graph structure design is simplistic, and the fixed adjacency matrix cannot adapt to the dynamic changes in health indicator associations, resulting in insufficient accuracy in capturing spatial associations; second, the feature-target association is disconnected, failing to integrate health indicators with SOH target information, leading to poor generalization performance with small samples; third, the fusion of spatiotemporal features is insufficient, as traditional time-series models use single-scale modeling, which cannot adapt to the heterogeneity of time-series features at different battery aging stages, resulting in poor capture of long-term time-series dependencies and local fluctuations. Summary of the Invention

[0005] To address the aforementioned problems in the prior art, this invention employs a lithium-ion battery health status prediction method and device based on a spatiotemporal graph neural network, which can dynamically capture spatial correlations, deeply fuse feature-target correlations, and accurately extract temporal features of the entire aging stage.

[0006] A method for predicting the state of health (SOH) of a lithium-ion battery based on a spatiotemporal graph neural network includes: predicting the battery's SOH using a trained and converged SOH prediction model based on the electrical characteristic data during the charging and discharging process of the lithium-ion battery; the trained and converged SOH prediction model is obtained in the following manner:

[0007] S1: Collect raw electrical characteristic data of lithium-ion battery during charging and discharging, preprocess the raw electrical characteristic data, extract the health indicator HI sequence that is strongly correlated with the battery health state SOH, and divide the health indicator sequence into training set and test set according to the proportion.

[0008] S2: Construct a dual dynamic graph structure based on the health indicator sequence of the training set, including two graph structures, feature association graph G1 and target association graph G2. Use KL divergence to minimize the feature space distance between G1 and G2, and extract the spatial features of the two graphs through graph attention network GAT; S3: Input the spatial features after the fusion of the two graphs into the aging state evaluator, and output the aging state features.

[0009] S4: The aging state features and the spatial features after the fusion of the two images are input into the multi-scale adaptive temporal convolutional network (TCN). Adaptive weights are assigned to the multi-scale convolutional branches of the TCN according to the aging state features. Temporal features are extracted from the fused features after the aging state features and the spatial features after the fusion of the two images according to the weights. The temporal features and spatial features are concatenated to obtain the spatiotemporal fusion features.

[0010] S5: Input the spatiotemporal fusion features into the SOH prediction model, set the model training hyperparameters, complete the model parameter training based on the training set, and verify the convergence through the test set to obtain the trained and converged SOH prediction model.

[0011] The present invention also provides an electronic device, including at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores a computer program executed by the at least one processor, the computer program being executed by the at least one processor to enable the at least one processor to perform the lithium-ion battery health status prediction method based on spatiotemporal graph neural network described above.

[0012] The present invention also provides a computer-readable storage medium storing computer instructions, which are used to cause a processor to execute the above-described method for predicting the health status of lithium-ion batteries based on a spatiotemporal graph neural network.

[0013] Compared with the prior art, the beneficial effects achieved by the present invention include:

[0014] This invention designs a dual dynamic graph architecture, overcoming the limitations of a single graph structure. By constructing a feature association graph G1 and a target association graph G2, the spatial correlation between health indicators and the target correlation between health indicators and SOH are captured respectively. At the same time, the adjacency matrix is ​​dynamically updated based on the battery aging state, overcoming the limitations of the traditional single fixed graph structure, realizing accurate dynamic capture of spatial correlation during battery aging, and improving the accuracy and robustness of spatial feature extraction.

[0015] This invention introduces KL divergence as a loss term. By minimizing the feature space distance extracted by the dual dynamic graphs, implicit fusion of features and targets is achieved, strengthening the correlation between feature extraction and SOH prediction tasks. This solves the problem of feature-target disconnect in existing models, significantly improves the generalization ability of the model in small sample scenarios, and can adapt to SOH prediction of different working conditions and different types of lithium-ion batteries.

[0016] This invention designs an aging state evaluator and a multi-scale adaptive TCN module. By normalizing the number of cycles, it achieves accurate identification of battery aging stages. A three-branch multi-scale dilated convolutional structure is designed to adapt to the heterogeneity of temporal features in the early, middle, and late aging stages of the battery. Adaptive weights are assigned to each branch based on aging state features, making the temporal modeling more closely resemble the actual aging mechanism of the battery. This achieves effective extraction and fusion of multi-scale temporal features across the entire aging stage, solving the problems of long-term temporal dependencies and capturing local fluctuations. It also overcomes the rigidity defect of traditional temporal models that "share parameters across all stages," making the temporal modeling more closely resemble the actual aging state of the battery. Attached Figure Description

[0017] Figure 1 This is a schematic diagram illustrating the process of obtaining a converged SOH prediction model for lithium-ion battery health state prediction based on a spatiotemporal graph neural network according to the present invention.

[0018] Figure 2 This is a schematic diagram of the structural framework for obtaining the converged SOH prediction model of lithium-ion battery health state prediction based on spatiotemporal graph neural network of the present invention.

[0019] Figure 3 This is a schematic diagram of the process for preprocessing raw electrical characteristic data according to an embodiment of the present invention;

[0020] Figure 4 This is a schematic diagram of the process of steps S2 to S5 of the present invention;

[0021] Figure 5 This is a schematic diagram of an embodiment of the electronic device structure of the present invention. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] This embodiment provides a method for predicting the health status of lithium-ion batteries based on spatiotemporal graph neural networks, including: using a trained and converged SOH prediction model to predict the battery health status based on the electrical characteristic data during the charging and discharging process of the lithium-ion battery.

[0024] The converged SOH prediction model was obtained in the following manner, as follows: Figure 1 , Figure 2 , Figure 4 As shown.

[0025] S1: Collect raw electrical characteristic data of lithium-ion batteries during charging and discharging. Preprocess the raw electrical characteristic data to extract health indicator HI sequences that are strongly correlated with the battery's state of health (SOH). Divide the health indicator sequences into a training set and an independent test set according to a certain ratio. Use cross-validation on the training set to complete model training and parameter optimization to improve the model's generalization and robustness. Figure 3 As shown. The ratio can be 7:3, 8:2, etc., and this invention does not impose any particular limitation.

[0026] The raw electrical characteristic data includes current, voltage, temperature and charge / discharge time data during the lithium-ion battery charging and discharging process, and the data sources collected cover public battery datasets such as CALCE and NASAPCoE.

[0027] The preprocessing of the raw electrical characteristic data includes removing abnormal cyclic data using the 3σ criterion, filling missing values ​​using linear interpolation, and normalizing the data using Z-score standardization.

[0028] The health indicator sequence consists of five core health factors extracted from the battery charging profile: HI1: area under the current curve during constant current CC charging; HI2: area under the current curve during constant current-constant voltage CC-CV charging; HI3: area under the voltage curve during constant current charging; HI4: constant current charging stage period; and HI5: area under the incremental capacity IC curve for a specific voltage range. Health factors with an absolute correlation coefficient greater than 0.8 with SOH were selected through Pearson correlation analysis to form the health indicator sequence.

[0029] S2: Construct a dual dynamic graph structure based on the health indicator sequence of the training set, including two graph structures, feature association graph G1 and target association graph G2. Use KL divergence to minimize the feature space distance between G1 and G2, and extract the spatial features of the two graphs through graph attention network GAT to obtain the spatial features after the fusion of the two graphs, thereby realizing the implicit fusion of features and targets.

[0030] S21: Construct the feature association graph G1;

[0031] ,

[0032] in, This is a set of nodes, where each node corresponds to one filtered health indicator. Let the boundary set represent the spatial relationships between health indicators. The weighted adjacency matrix of nodes is an undirected symmetric matrix, and its calculation formula is:

[0033] ,

[0034] in, Representing the characteristic sequence and characteristic sequences The weighted adjacency matrix, For characteristic sequences and characteristic sequences covariance, , They are the characteristic sequences. , The standard deviation reflects the degree of dispersion of the characteristic sequence;

[0035] S22: Construct the target association graph G2;

[0036] ,

[0037] in, This indicates updating the node set, in Based on the addition of SOH nodes, Let the boundary set represent the correlation between health indicators and SOH targets. The weighted adjacency matrix is ​​an undirected symmetric matrix. The similarity between nodes is calculated using the Pearson correlation coefficient between health indicators and SOH as the edge weight value.

[0038] S23: Dual-graph spatial feature extraction;

[0039] A two-layer graph attention network (GAT) is used to extract features from G1 and G2 respectively. Each hidden layer of GAT has a dimension of 64. A self-attention mechanism is introduced to automatically learn the weights of neighbor nodes, and the spatial features of G1 are output separately. Spatial characteristics of G2 ;

[0040] S24: Feature-target implicit fusion;

[0041] Using KL divergence as the loss term, minimize and The characteristic space distance, the KL divergence calculation formula is:

[0042] ,

[0043] Where X is the set of discrete sampling points in the feature space. For a single sampling point in the feature space, Spatial features extracted from feature association graph G1 The characteristic probability distribution, Spatial features extracted from feature association graph G2 The feature probability distribution is determined by minimizing the KL divergence, forcing the feature space distribution without SOH nodes to approximate the target space distribution containing SOH nodes. This allows the pure features extracted by G1 to indirectly carry SOH target information, achieving unsupervised deep fusion of features and targets. The fused output is a dual-graph spatial feature. D is the feature dimension.

[0044] S3: Input the spatial features after fusion of the two images into the aging state evaluator and output the aging state features, which are monotonically positively correlated with the degree of battery aging.

[0045] S31: Use monitoring signals Learning through supervised training The mapping relationship to battery aging level, outputting aging state characteristics. , The closer the value is to 0, the lower the degree of battery aging. The closer it is to 1, the higher the degree of battery aging.

[0046] The battery cycle number n is normalized and then used as a monitoring signal. The normalization formula is: In the formula, This represents the minimum number of battery cycles. This represents the maximum number of battery cycles. ;

[0047] The input to the aging condition assessor is the spatial features obtained by fusion of two maps. The aging state evaluator is a lightweight regression head structure, consisting of two fully connected layers, a ReLU activation function, and a Dropout layer. The hidden layer dimension of the first fully connected layer can be set to 64, the output layer dimension of the second fully connected layer can be set to 1, and the dropout rate of the Dropout layer can be 0.2.

[0048] S4: The aging state features and the spatial features after the fusion of the two images are input into the multi-scale adaptive temporal convolutional network (TCN). Adaptive weights are assigned to the multi-scale convolutional branches of the TCN according to the aging state features. Temporal features are extracted from the fused features after the aging state features and the spatial features after the fusion of the two images according to the weights. The temporal features and spatial features are concatenated to obtain the spatiotemporal fusion features.

[0049] S41: Based on the aging state characteristics, perform branch temporal feature extraction on the fused features obtained by fusing the aging state characteristics and the spatial features after fusing the two graphs.

[0050] In this embodiment, the TCN network is a three-branch multi-scale dilated convolutional structure, with each convolutional kernel having a size of k=3. The dilation coefficients of the three branches are respectively... =1、 =2、 =4, each branch contains dilated convolutional layers, weight normalization layers, ReLU activation function layers and dropout layers, with a dropout rate of 0.2 for the dropout layers;

[0051] In particular, if <0.3, coefficient of thermal expansion The fine-grained branch with a value of 1 extracts the aging state features and the spatial features after fusion of the two graphs, resulting in the high-frequency local temporal features of early battery aging, and outputs the temporal features. ;

[0052] If 0.3≤ ≤0.7 coefficient of thermal expansion The medium-granularity branch with a value of 2 extracts the aging state features and the spatial features after fusion of the two graphs, resulting in the mesoscale temporal features of the battery's mid-term aging, and outputs the temporal features. ;

[0053] like >0.7, coefficient of thermal expansion The coarse-grained branch with a value of 4 extracts the aging state features and the spatial features obtained by fusing the two graphs, resulting in low-frequency, long-time-series features of late-stage battery aging, and outputs the time-series features. ;

[0054] S42: Adaptive weight allocation of time-series features is performed using the Softmax function based on aging state characteristics.

[0055] Characteristics of aging The Softmax function maps these weights to network branch weights. The formula for the Softmax function is:

[0056] ,

[0057] in, This represents the weight coefficient of the network branch, whose value is automatically adjusted during model training, but must meet certain conditions. NC represents the number of network branches, and in this embodiment, NC is preferably 3.

[0058] S43, adaptive weight allocation is used to perform multi-scale temporal feature fusion on the branch temporal features to obtain the fused temporal features. :

[0059] .

[0060] S44: Concatenate spatial features with fused temporal features to obtain spatiotemporal fused features.

[0061] ,

[0062] in, This is the feature splicing function.

[0063] S5: Input the spatiotemporal fusion features into the SOH prediction model, set the model training hyperparameters, complete the model parameter training based on the training set, and verify the convergence through the test set to obtain the trained and converged SOH prediction model;

[0064] Using the spatiotemporal fusion feature F as input, feature mapping is completed through a fully connected layer, and the SOH prediction value is output. The formula is:

[0065] ,

[0066] In the formula, Let W be the mapping function of the fully connected layer, W be the weight matrix of the fully connected layer, and b be the bias term. y represents the predicted SOH value output by the model, corresponding to the actual SOH value.

[0067] The model training hyperparameters include: the optimizer is Adam, the learning rate is set to 0.001, the batch size is 32, the number of training epochs is 100, early stopping is used to prevent overfitting, and the patience value for early stopping is set to 10; mean squared error (MSE) is used as a partial loss function, and the formula for calculating MSE is:

[0068] ,

[0069] In the formula, M is the number of samples in the training set. The true SOH value for the i-th sample. Let be the predicted SOH value for the i-th sample.

[0070] The formula for the total loss function of the model is:

[0071] ,

[0072] In the formula, This indicates the SOH prediction error. It is a feature-target fusion constraint. It is a balance coefficient used to adjust the ratio of the two.

[0073] The feature-target fusion constraint is to force the feature association graph space features without SOH nodes to be fused. The probability distribution of the target association graph spatial features containing SOH nodes. By approximating the probability distribution of the target space and minimizing the distribution difference between the two feature spaces, implicit alignment between the feature space and the target space is achieved, enabling the pure feature space to indirectly carry SOH target information and completing the deep fusion of features and target.

[0074] The model convergence verification uses Mean Absolute Percentage Error (MAPE), Mean Absolute Error (MAE), and Root Mean Square Error (RMSE) as evaluation indicators, and the calculation formulas are as follows:

[0075] ,

[0076] ,

[0077] ,

[0078] In the formula, MAPE is used as the core indicator to determine the overall prediction accuracy of the model, MAE verifies the prediction stability of the whole sample, and RMSE tests the robustness of the model to abnormal samples. When the MAPE of the test set is less than 5%, the fluctuation of MAE and RMSE is less than 3%, and the loss function changes less than 1e-4 for 10 consecutive rounds, the model training is considered to have converged.

[0079] Figure 5 A schematic diagram of an electronic device that can be used to implement embodiments of the present invention is shown. The electronic device can be any form of digital computer or mobile computing device, such as a laptop computer, desktop computer, server, blade server, personal digital assistant, smartphone, and other similar computing device terminals.

[0080] The electronic device includes a processor 91, a memory 92, an input / output interface 93, a communication interface 94, and a bus 95. The processor 91, memory 92, input / output interface 93, and communication interface 94 are interconnected via the bus 95.

[0081] Processor 91 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities, such as a central processing unit (CPU), a graphics processing unit (GPU), a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), and / or other programmable logic devices. Processor 91 can control other components in an electronic device to perform desired functions.

[0082] Memory 92 is used to store computer programs that can run on the processor. Memory 92 may include volatile memory cells, such as random access memory (RAM) and / or cache memory; memory 92 may also include non-volatile memory cells, such as read-only memory (ROM), erasable programmable read-only memory (EPROM), and electrically erasable programmable read-only memory (EEPROM).

[0083] Input / output interface 93 is used for data exchange with external devices. Input devices may include keyboards, mice, touchscreens, microphones, cameras, etc.; output devices may include monitors, speakers, printers, etc. Communication interface 94 enables electronic devices to communicate with other devices or networks, and may include network interface cards, modems, wireless communication modules, etc. Bus 95 is used to connect the various components of electronic devices to realize the transmission of data and control signals.

[0084] The processor 91 executes computer programs and other data processing, such as implementing the steps of any of the above-described methods for predicting the health status of lithium-ion batteries based on spatiotemporal graph neural networks when executing the computer program. Specifically, the processor 91 reads the computer program from the memory 92 and executes the steps of any of the above-described methods for predicting the health status of lithium-ion batteries based on spatiotemporal graph neural networks.

[0085] In embodiments of the present invention, a computer-readable storage medium is also provided, which stores a computer program. When executed by a processor, the computer program implements the steps of any of the above-described methods for predicting the health status of lithium-ion batteries based on spatiotemporal graph neural networks. The computer-readable storage medium may include, but is not limited to, various tangible media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), portable compact disk read-only memory (CD-ROM), magnetic disks, or optical disks. The above embodiments further illustrate the objectives, technical solutions, and advantages of the present invention. It should be understood that the above embodiments are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, or improvements made to the present invention within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for predicting the health status of lithium-ion batteries based on spatiotemporal graph neural networks, comprising: A convergent State of Health (SOH) prediction model is used to predict the state of health of a lithium-ion battery based on its electrical characteristic data during charging and discharging. The convergent SOH prediction model is obtained in the following manner: S1: Collect raw electrical characteristic data of lithium-ion battery during charging and discharging, preprocess the raw electrical characteristic data, extract the health indicator HI sequence that is strongly correlated with the battery health state SOH, and divide the health indicator sequence into training set and test set according to the proportion. S2: Construct a dual dynamic graph structure based on the health indicator sequence of the training set, including two graph structures, feature association graph G1 and target association graph G2. Use KL divergence to minimize the feature space distance between G1 and G2, and extract the spatial features of the two graphs through graph attention network GAT. S3: Input the spatial features after fusion of the two images into the aging state evaluator and output the aging state features; S4: The spatial features after the two images are fused with the aging state features, and the fused features are then input into a multi-scale adaptive temporal convolutional network (TCN). Based on the aging state features, adaptive weights are assigned to the multi-scale convolutional branches of the TCN, and temporal features are extracted based on these weights. Finally, the extracted temporal features are concatenated with the spatial features to obtain the spatiotemporal fusion features. S5: Input the spatiotemporal fusion features into the SOH prediction model, set the model training hyperparameters, complete the model parameter training based on the training set, and verify the convergence through the test set to obtain the trained and converged SOH prediction model.

2. The method according to claim 1, characterized in that, The raw electrical characteristic data mentioned in S1 includes current, voltage, temperature and charge / discharge time data during the charging and discharging process of lithium-ion batteries. The data sources collected cover public battery datasets such as CALCE and NASAPCoE.

3. The method according to claim 1, characterized in that, The preprocessing of the raw electrical characteristic data described in S1 includes removing abnormal cyclic data using the 3σ criterion, filling missing values ​​using linear interpolation, and normalizing the data using Z-score standardization.

4. The method according to claim 1, characterized in that, The health indicator sequence described in S1 consists of five core health factors extracted from the battery charging profile: HI1: area under the current curve during constant current CC charging; HI2: area under the current curve during constant current-constant voltage CC-CV charging; HI3: area under the voltage curve during constant current charging; HI4: constant current charging stage period; and HI5: area under the incremental capacity IC curve for a specific voltage range. Health factors with an absolute correlation coefficient greater than 0.8 with SOH were selected through Pearson correlation analysis to form the health indicator sequence.

5. The method according to claim 1, characterized in that, S2 include: S21: Construct the feature association graph G1; ， in, This is a set of nodes, where each node corresponds to one filtered health indicator. Let the boundary set represent the spatial relationships between health indicators. The weighted adjacency matrix of nodes is an undirected symmetric matrix, and its calculation formula is: ， in, Representing the characteristic sequence and characteristic sequences The weighted adjacency matrix, For characteristic sequences and characteristic sequences covariance, , They are the characteristic sequences. , The standard deviation reflects the degree of dispersion of the characteristic sequence; S22: Construct the target association graph G2; ， in, To update the node set, in Based on the addition of SOH nodes, Let the boundary set represent the correlation between health indicators and SOH targets. For a weighted adjacency matrix, the Pearson correlation coefficient between health indicators and SOH is used to calculate the similarity between nodes, which is then used as the edge weight value. It is an undirected symmetric matrix; S23: Dual-graph spatial feature extraction; A two-layer graph attention network (GAT) is used to extract features from G1 and G2 respectively. Each hidden layer of GAT has a dimension of 64. A self-attention mechanism is introduced to automatically learn the weights of neighbor nodes, and the spatial features of G1 are output separately. Spatial characteristics of G2 ; S24: Feature-target implicit fusion, outputting dual-graph space features. Using KL divergence as the loss term, minimize and The characteristic space distance, the KL divergence calculation formula is: ， Where X is the set of discrete sampling points in the feature space. For a single sampling point in the feature space, Spatial features extracted from feature association graph G1 The characteristic probability distribution, Spatial features extracted from feature association graph G2 The characteristic probability distribution; The output is a dual-graph spatial feature after fusion, where D is the feature dimension.

6. The method according to claim 1, characterized in that, S3 include: Using monitoring signals Learning through supervised training The mapping relationship to battery aging level, outputting aging state characteristics. ; The battery cycle number n is normalized and then used as a monitoring signal. The normalization formula is: In the formula, This represents the minimum number of battery cycles. This represents the maximum number of battery cycles.

7. The method according to claim 1, characterized in that, S4 include: S41: Extract branch time series features based on aging state characteristics; S42: Adaptive weight allocation of aging state features is performed using the Softmax function; S43: Adaptive weight allocation is used to fuse the branch time series features at multiple scales to obtain the fused time series features; S44: Spatial features are concatenated with fused temporal features to obtain spatiotemporal fused features.

8. The method according to claim 1, characterized in that, S41 includes: If aging characteristics <0.3, fine-grained branching extracts high-frequency local temporal features of aging state features and spatial features after fusion of two graphs; If 0.3≤ ≤0.7, medium-granularity branch extracts the aging state features and the spatial features after fusion of the two graphs, and the mesoscale temporal features of the fused features; like >0.7, coarse-grained branching extracts low-frequency long-time features of aging state features and spatial features after fusion of two graphs.

9. An electronic device, comprising at least one processor; and a memory communicatively connected to said at least one processor; wherein, The memory stores a computer program that is executed by the at least one processor, wherein the at least one processor is capable of executing the lithium-ion battery health status prediction method based on spatiotemporal graph neural network as described in any one of claims 1-8.

10. A computer-readable storage medium storing computer instructions, characterized in that, The computer instructions are used to cause the processor to implement the lithium-ion battery health status prediction method based on spatiotemporal graph neural network as described in any one of claims 1-8.

Images