CNN and GNN-based battery specific energy prediction method and system

Abstract

The invention provides a CNN and GNN-based battery specific energy prediction method and system, and the method comprises the steps: inputting non-graph structure data into a trained CNN model, and obtaining local features; inputting the graph structure data into the trained GNN model to obtain global features; fusing the local features and the global features to obtain fused features; and inputting the fusion features into a trained prediction model to obtain a battery specific energy prediction value. According to the invention, high-precision prediction of the specific energy of the battery is realized.

Description

Technical Field

[0001] The present invention belongs to the technical field of battery performance prediction, and in particular relates to a battery specific energy prediction method and system based on CNN and GNN. Background Art

[0002] Battery specific energy, a core performance metric, refers to the amount of electrical energy a battery can store per unit weight or volume. This energy directly impacts battery performance in various applications, such as electric vehicles, smart devices, and wearables. With the continuous advancement of new energy technologies, improving battery specific energy, particularly through rational battery design and material optimization, has become a key focus in battery research and development. Therefore, accurately and efficiently predicting battery specific energy has become a key issue in battery design and performance optimization.

[0003] Traditionally, battery specific energy prediction methods rely on the accumulation of experimental data and traditional machine learning algorithms, such as support vector machines (SVM) and random forests (RF). These methods estimate specific energy by learning the relationship between input features and battery performance from experimental data. However, the prediction accuracy of traditional methods is often constrained by the following factors: 1. Limitations of experimental data: Traditional experimental data requires a large number of samples and often has limitations in efficiency and accuracy, especially in the early stages of battery design, when data is scarce and unrepresentative. 2. Feature selection and high-dimensional data issues: Traditional machine learning algorithms are prone to the "curse of dimensionality" problem when processing high-dimensional data, especially when the data comes from different sources (such as images, molecular structures, etc.). How to efficiently extract effective features to improve prediction accuracy is a challenge that needs to be addressed. 3. Poor model generalization ability: Most traditional methods cannot effectively adapt to new and unknown battery designs and material combinations, resulting in poor generalization ability of the model in practical applications and difficulty in predicting the performance of different types of batteries.

[0004] Furthermore, traditional prediction methods often simplify the task of predicting the specific energy of battery materials to a single-dimensional feature analysis, relying solely on the ratio of chemical composition or local features of the crystal structure, while ignoring the complex relationships and multi-level influencing factors. Consequently, these methods often suffer from large errors and low accuracy when faced with complex battery material data. Summary of the Invention

[0005] The purpose of the present invention is to provide a battery specific energy prediction method and system based on CNN and GNN, combining the advantages of CNN and GNN to achieve high-precision prediction of battery specific energy. The technical solution adopted is:

[0006] A battery specific energy prediction method based on CNN and GNN includes the following steps:

[0007] Input non-graph structure data into the trained CNN model to obtain local features;

[0008] Input graph structure data into the trained GNN model to obtain global features;

[0009] The fusion feature is obtained by fusing the local feature and the global feature;

[0010] The fused features are input into the trained prediction model to obtain the battery specific energy prediction value.

[0011] Preferably, the non-graph structured data includes:

[0012] Chemical elements of the positive electrode material and the ratio of chemical elements of the positive electrode material;

[0013] Chemical elements of the negative electrode material and the ratio of chemical elements of the negative electrode material.

[0014] Preferably, the graph structure data includes: node features and edge features.

[0015] Preferably, the image structure data is acquired based on a microscopic image of the battery.

[0016] Preferably, a concatenation or weighted summation method is used to obtain fusion features.

[0017] Preferably, the prediction model includes at least one fully connected layer.

[0018] Preferably, the CNN model includes: a convolutional layer and a pooling layer connected in sequence.

[0019] Preferably, the GNN model includes: a graph convolution layer and a graph pooling layer connected in sequence.

[0020] A battery specific energy prediction system based on CNN and GNN, including:

[0021] CNN model, used to extract local features from non-graph structured data;

[0022] GNN model, used to extract global features from graph structure data;

[0023] Fully connected layer, used to fuse local features and global features to generate fused features;

[0024] and prediction model, whose input is connected to the output of the fully connected layer, which is used to map the fusion features to the battery specific energy.

[0025] First, a CNN is used to deeply process non-graph structure data, accurately extracting its spatial features. Simultaneously, a GNN is used to efficiently process crystal structure data, precisely capturing its graph structure features. Second, the features extracted by the CNN and GNN are deeply fused to construct a unified feature vector. Finally, a fully connected layer is used to accurately map this feature vector to the predicted battery specific energy. Furthermore, advanced techniques such as cross-validation and regularization are employed to optimize the model, further improving prediction accuracy and generalization capabilities.

[0026] Compared with the prior art, the advantages of the present invention are:

[0027] 1. Improve prediction accuracy: By combining the advantages of CNN and GNN, full use is made of the battery's non-image data (microstructure image data) and image data (molecular or crystal structure data) to achieve high-precision prediction of battery specific energy.

[0028] 2. Reduce computational complexity:

[0029] By optimizing network structure and algorithms, we reduce computational complexity and improve prediction efficiency. We enhance model generalization capabilities: Through data augmentation, regularization techniques, dropout mechanisms, and other means, we enhance the generalization capabilities of the model and improve the stability and reliability of prediction results. We promote the development of battery technology: We provide strong support for battery research and development, production, and application, and promote the continued development and advancement of battery technology.

[0030] 3. Multimodal data processing:

[0031] It supports multimodal data processing and can simultaneously process data from different sources, such as microscopic structural data of battery materials (such as SEM and TEM images) and molecular or crystal structure data. Image data and structural data typically have different feature representation methods and processing requirements. By using a model that combines CNN and GNN, it can effectively extract useful information from both types of data and integrate it through a feature fusion module, providing a more comprehensive and rich feature representation for battery specific energy prediction.

[0032] This multimodal data processing approach not only improves the accuracy of battery specific energy predictions but also supports comprehensive evaluation of battery material properties, including stability, conductivity, energy density, and other characteristics. This technology has broad application prospects in battery material research, development, and optimization.

[0033] In order to verify the accuracy and reliability of the proposed battery specific energy prediction method, the proposed model was evaluated through extensive experiments.

[0034] By comparing with traditional methods (such as regression analysis, support vector machine, etc.), the advantages of the method of the present invention in prediction accuracy, computational efficiency and generalization ability are verified.

[0035] Experimental results demonstrate that the deep learning model combining CNN and GNN significantly improves the accuracy of battery specific energy prediction and is capable of processing more complex multidimensional data. Furthermore, through evaluation techniques such as cross-validation and leave-one-out, the proposed method maintains good generalization across diverse datasets. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 Flowchart of the battery specific energy prediction method based on CNN and GNN;

[0037] Figure 2 This is the CNN model structure diagram;

[0038] Figure 3 It is the structure diagram of the GNN model;

[0039] Figure 4 Generate schematics for fused features. DETAILED DESCRIPTION

[0040] The following is a more detailed description of the CNN- and GNN-based battery specific energy prediction method and system of the present invention, with reference to schematic diagrams. These schematic diagrams illustrate preferred embodiments of the present invention. It should be understood that those skilled in the art may modify the present invention described herein while still achieving the beneficial effects of the present invention. Therefore, the following description should be understood as a general guide for those skilled in the art and not as a limitation of the present invention.

[0041] like Figures 1 to 4 , a battery specific energy prediction method based on CNN and GNN, including:

[0042] Step 1: Data preparation and preprocessing.

[0043] Step 1A: Collect non-graph structure data of batteries from the Materials project database and divide it into training set t1 and training set T1.

[0044] Non-graph structured data includes:

[0045] Chemical elements of positive electrode materials, chemical element ratios of positive electrode materials; positive electrode materials such as LiCoO2, LiFePO4, etc.

[0046] Chemical elements of the negative electrode material, chemical element ratio of the negative electrode material; negative electrode materials such as graphite, silicon-carbon composite materials, etc.).

[0047] The chemical composition of the electrolyte (such as carbonate solvent and lithium salt LiPF6, etc.).

[0048] The separator composition (such as polyethylene, polypropylene), for example, the electrode positive electrode material (the ratio of nickel, manganese, and cobalt (Ni:Mn:Co).

[0049] The capacity decay of the battery occurs as the number of charge and discharge cycles increases. Commonly used capacity decay models include: C(t)=C0-αt.

[0050] C(t) is the effective capacity of the battery at time t, C0 is the initial capacity, and α is the decay efficiency.

[0051] Figure 1 The data acquisition module refers to the official API of the materials project library.

[0052] Step 1B: Collect the graph structure data of batteries from the Materials project database and divide it into training set t2 and training set T2.

[0053] Graph structure data includes atomic coordinates and chemical bond information.

[0054] Graph structure data refers to the molecular structure data or crystal structure data of the battery.

[0055] In this embodiment, molecular structure data of batteries is used.

[0056] The molecular structure diagram composed of molecular structure data is represented as follows:

[0057] G=(V,E)V={vi}E={(vi,vj,ωj)}

[0058] Where V is the node set; E is the edge set; ωj is the edge weight;

[0059] vi in the node set is the i-th atom.

[0060] In the edge set:

[0061] vj and vi are two atoms (nodes) connected by chemical bonds;

[0062] ωj represents the weight of the edge, which usually contains the following information:

[0063] The type of bond (single bond, double bond, triple bond, etc.) can be represented by a numerical value (such as 1, 2, 3).

[0064] The length of the bond (bond length).

[0065] Bond strength or other physicochemical properties (such as bond energy, bond order, etc.).

[0066] like Figure 1As shown, the input node features refer to the node set.

[0067] The input edge features refer to edge sets.

[0068] Among them, the number of data in training set T2 is equal to that in training set T1.

[0069] When constructing the input data format, the chemical composition of the battery material (for example, "LiCoO2" is converted to "Lithium Cobalt Oxide"), crystal structure, and other information are converted into text format and mapped one-to-one with the battery voltage data as input to the XLM-R model. In other words, the XLMR model is used to encode the chemical composition and voltage data in the Materials Project database.

[0070] To provide effective model output, the XLMR model requires accurate and logical data as input to learn the correct patterns. The one-to-one correspondence between chemical composition and voltage data enables the model to establish a mapping between material, voltage, and specific energy, facilitating specific energy calculations.

[0071] Step 1C, pretreatment.

[0072] Denoising, contrast enhancement, image cropping, and converting molecular or crystal structure data into graph-structured representations.

[0073] The reason for converting to graph structure representation is to make it suitable for GNN input by encoding the information of atomic nodes and chemical bond edges.

[0074] Data preprocessing also includes the following operations:

[0075] Data cleaning and missing value processing are performed to ensure the consistency and completeness of the dataset.

[0076] Perform data augmentation operations such as scaling, rotating, and cropping on image data to increase the diversity of input images to adapt to the size requirements of network input; at the same time, perform standardized encoding on molecular structure data (such as converting chemical bonds and atomic properties into numerical features) to facilitate the network to effectively understand and process its complex topological structure.

[0077] Step 2: Build a CNN model and obtain the local feature set.

[0078] Step 2A: Design the CNN model.

[0079] CNN model is convolutional neural network model.

[0080] The CNN model contains at least one convolutional layer, at least one pooling layer, at least one activation function layer, and at least one batch normalization layer. Figure 2In the figure, the activation function layer and batch normalization layer are not shown.

[0081] Convolution feature extraction formula: fconv(X)=σ(W*X+b)

[0082] Graph convolution formula:

[0083] Where W-convolution kernel; b-bias; σ-activation function;

[0084] -The l-th layer node features of node i;

[0085] N(i)-the set of neighbors of node i;

[0086] W (l) -weight matrix;

[0087] X - input data.

[0088] The convolution operation can be expressed as follows:

[0089]

[0090] x-input image or feature map;

[0091] ω-convolution kernel (filter);

[0092] y-output feature map;

[0093] (i,j) - position index, index i has a different meaning from node i;

[0094] m,n - the size of the convolution kernel.

[0095] The operations in a CNN module can be expressed as a combination of convolution and activation functions:

[0096] F l+1 =f(W l *F l +b l )

[0097] F l -Input feature map of layer l;

[0098] W l -Convolution kernel weights;

[0099] *-convolution operation;

[0100] b l - bias term;

[0101] f - activation function.

[0102] Step 2B: Use the preprocessed training set t1 and the feature set corresponding to the training set t1 to train the CNN model (supervised learning) so that it can automatically learn the features in the image.

[0103] Layer attention mechanism formula:

[0104] α ij - attention weight; a-weight vector.

[0105] Step 2C: Input the feature set corresponding to the preprocessed training set T1 into the trained CNN model to obtain the local feature set.

[0106] The CNN model also includes data augmentation techniques such as rotation, scaling, and flipping to increase data diversity and prevent overfitting.

[0107] Step 3: Build a GNN model and obtain the global feature set.

[0108] Step 3A: Design the GNN model.

[0109] GNN model, that is, graph neural network model.

[0110] The GNN model contains at least one graph convolution layer, at least one graph attention layer, and at least one graph pooling layer.

[0111] The core idea of GNN is to update the representation of nodes through message passing between nodes. Suppose a graph G = (V, E), where V is the set of nodes and E is the set of edges.

[0112] During the message passing process of a graph neural network, the representation of a node is updated based on the representation of its neighboring nodes. The common update formula is as follows:

[0113]

[0114] in, - the representation of node i at layer k;

[0115] N(i)-the set of neighbor nodes of node i;

[0116] W (k) - The trainable weight matrix of layer k;

[0117] σ - non-linear activation function.

[0118] The graph convolution formula is as follows:

[0119]

[0120] in - normalized adjacency matrix;

[0121] X-node feature matrix, size N x F (N is the number of nodes, F is the number of features);

[0122] W - learnable weight matrix;

[0123] σ - non-linear activation function.

[0124] Step 3B: Use the preprocessed training set t2 and the feature set corresponding to the training set t2 to train the CNN model so that it can process complex graph structure data and capture the complex relationship between atoms and chemical bonds.

[0125] Step 3C: Input the preprocessed training set T2 into the trained GNN model to obtain the global feature set.

[0126] The GNN model also includes graph embedding technology, which is used to embed high-dimensional graph structure data into a low-dimensional vector space for subsequent processing.

[0127] Step 4: Based on the local feature set and the global feature set, a fused feature dataset containing several fused features is obtained.

[0128] Use methods such as splicing or weighted summation to fuse the features extracted by CNN and GNN to form a fused feature vector;

[0129] Feature fusion formula:

[0130] Splicing method: z = Concat(z CNN ,z GNN )

[0131] Weighted summation method: z = α*z CNN +(1-α)*z GNN

[0132] α-weighting coefficient;

[0133] Splicing:

[0134] F fusion =[F CNN ; F GNN ]

[0135] Weighted sum:

[0136] F fusion =aF CNN +bF GNN

[0137] Among them, F CNN and F GNN are the features extracted by CNN and GNN respectively, and a and b are weight fusion.

[0138] PCA dimensionality reduction formula:

[0139] Z reduced =Z*W PCA

[0140] W PCA - Dimensionality reduction matrix for principal component analysis.

[0141] Among them, feature fusion also includes feature selection and dimensionality reduction processing to reduce the number and redundancy of features and improve the generalization ability of the model.

[0142] Step 5: Build the prediction module.

[0143] Step 5A: Design a prediction model containing one or more fully connected layers.

[0144] Step 5B: Use the fused feature set to train the prediction model so that it can map the feature vector to the output space of battery specific energy.

[0145] Among them, the battery specific energy dataset used for training is obtained from an open source database.

[0146] Fully connected layer mapping formula: y = σ (W out z+b out )

[0147] W out -weight matrix; b out -bias; σ-activation function; z-fusion feature;

[0148] The output of the fully connected layer can be expressed as: y = σ(WF fusion +b)

[0149] Among them, y-prediction result, W-weight matrix, b-bias term.

[0150] The prediction model can learn the mapping relationship between the feature vector and the battery specific energy, and achieve high-precision prediction of the battery specific energy.

[0151] The prediction model also includes other technologies in deep learning, such as residual network (ResNet), dense connection network (DenseNet), etc., to further improve the prediction accuracy.

[0152] Step 6: Input the preprocessed non-graph structure data and graph structure data of another battery into the trained CNN model and GNN model to obtain the predicted value of the battery specific energy.

[0153] Input non-graph structure data into the trained CNN model to obtain local features;

[0154] Input graph structure data into the trained GNN model to obtain global features;

[0155] The fusion feature is obtained by fusing the local feature and the global feature;

[0156] The fused features are input into the trained prediction model to obtain the battery specific energy prediction value.

[0157] The CNN module focuses on extracting local features of image data, and uses components such as convolutional layers and pooling layers to extract detailed features of the battery microstructure.

[0158] The convolution operation of CNN automatically learns and extracts the spatial features between different chemical elements, and can capture the underlying patterns of chemical composition combinations. This enables the system to identify the impact of different chemical composition combinations on comparative energy, improving the accurate understanding of materials and enhancing the accuracy of predictions.

[0159] The GNN module focuses on the global features of molecular or crystal structures, processing information about the relationships between atoms and chemical bonds through graph convolutional layers and graph attention layers. This multimodal feature extraction method effectively improves the model's prediction accuracy for battery performance.

[0160] After CNN and GNN feature extraction, the feature fusion module fuses the two types of features to form a composite feature vector containing image information and structural information.

[0161] The fully connected layer linearly transforms the input features using a weight matrix to generate specific energy predictions. To improve prediction accuracy, this module can also use deep learning techniques such as residual connections and dense connections to reduce the vanishing gradient problem during training and accelerate convergence.

[0162] Specific energy is usually an important performance indicator of batteries. It indicates the energy that can be stored per unit mass or volume of the battery.

[0163]

[0164] Where: E space -Specific energy (unit: Wh / kg or Wh / L)

[0165] E total -Total battery energy (unit: Wh)

[0166] m-battery mass or volume (unit: kg or L)

[0167] In battery specific energy prediction, regression models are usually used for prediction.

[0168] A common choice of loss function is the mean squared error (MSE):

[0169]

[0170] Where: n is the number of samples; y i is the true value; is the predicted value.

[0171] Step 7: Post-processing and optimization.

[0172] Smoothing or denoising the prediction results to improve the stability and reliability of the prediction;

[0173] Smoothing formula:

[0174] N: The size of the smoothing window

[0175] Optimize the model based on the evaluation results, including adjusting the network structure, increasing the amount of training data, using regularization techniques, etc.

[0176] In summary, the prediction results are post-processed, including smoothing and denoising, to improve the stability and reliability of the predictions. Simultaneously, the model is optimized based on the evaluation results, including adjusting the network structure, increasing the amount of training data, and using regularization techniques (such as dropout) to further improve prediction accuracy and stability.

[0177] Smoothing methods such as moving average and exponential smoothing are used to eliminate noise and fluctuations in the forecast results. Denoising methods such as Gaussian filtering and median filtering are used to effectively remove outliers and improve the stability of the forecast results.

[0178] Adversarial sample testing aims to verify the robustness of the model under interference.

[0179] x ′ =x+δThe adversarial sample x' consists of the original sample x and the perturbation δ.

[0180] In addition, the evaluation indicators involved in the training process include MAE, MSE, R 2 The coefficient of determination, among other factors, is used to comprehensively evaluate the model's predictive accuracy. Cross-validation and the leave-one-out method are used to ensure the model's generalization and stability. During training, regularization techniques (such as L2 regularization) are used to reduce overfitting, address underfitting, and improve the model's predictive accuracy.

[0181] This invention uses transfer learning and AutoML technologies to continuously optimize the model based on new battery data. Reinforcement learning and ensemble learning enhance the model's predictive power and stability. Ensemble learning further improves accuracy by fusing multiple models, reducing the bias of individual models.

[0182] This battery specific energy prediction system can be applied not only to the battery field but also to cross-domain data integration and application, promoting intelligent prediction work in other fields. Specific application scenarios include semiconductor material prediction, photovoltaic material research, catalyst performance prediction, and composite material performance evaluation.

[0183] For the prediction of semiconductor material properties, such as conductivity and carrier mobility, the following relationship can be used: σ = n*e*μ

[0184] Where: σ is the conductivity (S / m), n is the carrier concentration, e is the basic current, and μ is the carrier mobility

[0185] In the performance evaluation of composite materials, the focus is usually on the tensile strength, hardness, elastic modulus, etc. The elastic modulus of the composite material (Ecomposite) can be estimated using the regular mixture theory:

[0186]

[0187] Where: E composite - elastic modulus of the composite material; V f and V m are the volume fractions of fiber and matrix, respectively; E f and E m are the elastic moduli of the fiber and the collective, respectively

[0188] The system can predict semiconductor material properties such as conductivity and carrier mobility based on the material's microstructure and crystal structure. This technology is of great significance in the design of new semiconductor materials, providing data support for material modification and optimization, and promoting the development of the semiconductor industry.

[0189] The technical framework of this embodiment is highly scalable and can be applied to the performance prediction of other types of materials, such as new energy materials, semiconductor materials, and catalyst materials. By adjusting the input features and target variables of the model, the properties of different materials can be predicted.

[0190] In practical applications, the prediction method provided in this embodiment uses an automated training process for deep learning models and machine learning algorithms, which can be automatically updated and optimized based on new experimental data and material data (not data in the database). As new data is added, the model can continue to learn and optimize, ensuring that its prediction performance continues to improve.

[0191] The logic and / or steps represented in the flowchart or otherwise described herein may be considered, for example, as an ordered list of executable instructions for implementing logical functions, and may be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a system including a processor, or other system that can fetch and execute instructions from an instruction execution system, apparatus, or device).

[0192] The above description is merely a preferred embodiment of the present invention and does not limit the present invention in any way. Any person skilled in the art who, without departing from the scope of the present invention, makes any equivalent substitution, modification, or other changes to the technical solution and technical content disclosed in the present invention shall be deemed to be within the scope of the present invention and still fall within the scope of protection of the present invention.

Claims

1. A battery specific energy prediction method based on CNN and GNN, characterized in that: The following steps are involved: Input non-graph structure data into the trained CNN model to obtain local features; Input graph structure data into the trained GNN model to obtain global features; The fusion feature is obtained by fusing the local feature and the global feature; The fused features are input into the trained prediction model to obtain the battery specific energy prediction value.

2. The battery specific energy prediction method based on CNN and GNN according to claim 1, characterized in that: The non-graph structure data includes: Chemical elements of the positive electrode material and the ratio of chemical elements of the positive electrode material; Chemical elements of the negative electrode material and the ratio of chemical elements of the negative electrode material.

3. The battery specific energy prediction method based on CNN and GNN according to claim 1, characterized in that: The graph structure data includes: node features and edge features.

4. The battery specific energy prediction method based on CNN and GNN according to claim 3, characterized in that: The image structure data is obtained based on a microscopic image of the battery.

5. The battery specific energy prediction method based on CNN and GNN according to claim 1, characterized in that: Use splicing or weighted summation method to obtain fusion features.

6. The battery specific energy prediction method based on CNN and GNN according to claim 1, characterized in that: The prediction model includes at least one fully connected layer.

7. The battery specific energy prediction method based on CNN and GNN according to claim 1, characterized in that: The CNN model includes: a convolutional layer and a pooling layer connected in sequence.

8. The battery specific energy prediction method based on CNN and GNN according to claim 1, characterized in that: The GNN model includes: a graph convolution layer and a graph pooling layer connected in sequence.

9. A battery specific energy prediction system based on CNN and GNN, characterized in that: include: CNN model, used to extract local features from non-graph structured data; GNN model, used to extract global features from graph structure data; Fully connected layer, used to fuse local features and global features to generate fused features; and prediction model, whose input is connected to the output of the fully connected layer, which is used to map the fusion features to the battery specific energy.

Images