A lithium-ion battery health status prediction method based on graph neural network

Abstract

The present application discloses a method for predicting the health status of lithium-ion batteries based on graph neural networks. The method comprises the following steps: obtaining the health status of each half-cycle of the lithium-ion battery; constructing the nodes, features and labels of the nodes of the graph, and setting the window size; dividing the graph data into a training set, a validation set and a test set, and normalizing the graph data; using the graph neural network method to train and learn the training set data, and testing the test set samples, thereby realizing the prediction of the health status of the lithium-ion battery. The present invention can comprehensively consider and take into account the impact of historical charging and discharging on future health status, so the accuracy of the present invention is relatively higher; in addition, the present invention avoids the complicated feature engineering work of voltage, current and temperature.

Description

Technical Field

[0001] The present invention relates to the field of battery health status management, and in particular to a lithium-ion battery health status prediction method based on graph neural network. Background Art

[0002] In recent years, lithium-ion batteries have garnered industry recognition and policy support due to their high energy density, low self-discharge rate, high efficiency and stability, and long cycle life. They are widely used in aerospace, automotive, and power systems. However, frequent battery fire incidents both domestically and internationally have heightened awareness of the severity of the problem, placing battery safety high on the agenda. Battery safety is closely linked to aging, which corresponds to different states of health. Therefore, assessing the health of lithium-ion batteries can aid in ensuring safe and reliable operation. For the assessment of the health status of lithium-ion batteries, there are usually methods such as mechanism models, empirical models and data-driven methods. Among them, the mechanism model uses a series of partial differential equations to describe the health status of the battery. The model is accurate and highly interpretable, but the problem is that, on the one hand, the electrochemical parameters appearing in the equation need to be measured by precise instruments, the equipment is expensive, and it is difficult to carry out online; on the other hand, solving partial differential equations is more difficult, it is difficult to obtain analytical expressions, and the amount of calculation is large, which has certain requirements on computer performance; for empirical models, it refers to the use of formulas to fit the health status of the battery. This method has poor generalization performance and low accuracy; as for data-driven, it does not require modeling, but only needs to obtain the law of battery aging through mining and learning of historical data, and then complete the health status assessment. However, current battery health status assessments mainly focus on feature engineering for a certain process, stage, or segment of charging and discharging to extract health factors. For example, the paper "Lithium-ion Battery Health Status Estimation Based on Constant-Amplitude Charging Time" extracts constant-amplitude charging time as a feature for health status estimation during the constant-current charging phase of the charging process. However, the above-mentioned health status assessment is always ex post facto. If the health status of an unknown process can be predicted in advance, it will be of great benefit to the estimation of the state of charge and the research and judgment of life prediction. However, it is difficult to predict an unknown health status based solely on the previous health status or health factor. When it is necessary to use multiple health statuses or health factors before the health status to be predicted to predict the future, the sample may be in a non-Euclidean space, which will fail for most algorithms based on Euclidean space. Summary of the Invention

[0003] To solve the above problems, a method for predicting the health status of lithium-ion batteries based on graph neural networks is proposed. First, the duration of different charging and discharging processes or stages and the charging or discharging capacity of the corresponding half-cycle are collected; secondly, the points and edges are clarified, the attributes and labels of the nodes are determined, and the window size is set to form a graph; the graph data is divided into a training set and a test set, and the data is normalized; then, the training set data is trained and learned using the graph neural network method, and the test set samples are tested, thereby realizing the prediction of the health status of lithium-ion batteries. The graph neural network adopted in the present invention can be applied to both Euclidean space and non-Euclidean space, has a wide range of applicability, and has a broad range of application scenarios.

[0004] The purpose of the present invention is achieved by at least one of the following technical solutions.

[0005] A method for predicting the health status of lithium-ion batteries based on a graph neural network includes the following steps:

[0006] S1. Obtain the health status of each half cycle of the lithium-ion battery;

[0007] S2, construct the nodes, node features and labels of the graph, and set the window size;

[0008] S3. Divide the graph data into training set, validation set, and test set, and normalize the graph data;

[0009] S4. Use the graph neural network method to train and learn the training set data, and test the test set samples to achieve the health status prediction of lithium-ion batteries.

[0010] Furthermore, in step S1, the charging and discharging process includes a charging process or a discharging process; the charging and discharging stage includes a constant current charging stage, a constant voltage charging stage or a constant current discharging stage; the charging process includes a constant current charging stage and a constant voltage charging stage, and the discharging process includes a constant current discharging stage;

[0011] Completing one charging process and one discharging process is a complete charge and discharge cycle, while a half cycle means completing one charging process or one discharging process alone.

[0012] The health status is the ratio of the charging capacity or discharging capacity to the rated capacity. The rated capacity is the factory parameter of the battery. It should be noted that the discharge capacity is a negative value in the present invention.

[0013] Furthermore, in step S2, a complete charge-discharge cycle is regarded as a node, the health status of two adjacent half cycles is used as the node feature, and the health status of the half cycle after several charge-discharge cycles is used as the graph label; the number of nodes N, i.e., the window size, is set to establish the graph g, and M isomorphic graphs g are formed by sliding the window. m, m=1~M, graph set G=g1,g2,g3,...g M .

[0014] Furthermore, in step S3, M isomorphic graphs g m Divide into training set G train =g1 train , g2 train , g3 train ,...g p train , where the subscripts train and subscripts represent the training set, p represents the number of samples in the training set, and g p train Represents the training set G train The p-th training graph in the validation set G valid =g1 valid , g2 valid , g3 valid ,...g b valid , where the subscripts valid and valid represent the validation set, b represents the number of samples in the training set, and g b valid Represents the training set G valid The b-th training graph in the test set G test =g1 teat , g2 teat , g3 test ,...g q test , where the subscripts test and g represent the number of samples in the test set. q test Denotes the test set G test The g-th test image in , where p+b+q=M.

[0015] Furthermore, for the training set G train The training graph g in p train , find all training graphs g p train The maximum and minimum values ​​of the health status of the two half cycles (distinguished by f and a) in all nodes;

[0016] For the training set G train All nodes in all training graphs are normalized to extreme values:

[0017]

[0018]

[0019]

[0020] For the validation set G valid All validation graphs and test set G test All nodes in all test graphs are normalized to their extreme values:

[0021]

[0022]

[0023]

[0024] Among them, SOH f·s·t 、SOH a·s·t They represent the health status of the two half cycles before and after normalization of the t-th node in the s-th training graph in the training set; SOH f·j·k 、SOH a·j·k They represent the constant current charging time, constant voltage charging time, and constant current discharging time of the j-th verification graph or the k-th node in the test graph before normalization; SOH f·s·t 、SOH a·s·t is the normalized constant current charging time, constant voltage charging time, and constant current discharging time of the tth node in the sth training graph in the training set; SOH f·j·k 、SOH a·j·k SOH is the health status of the two half-cycles before and after normalization of the k-th node in the j-th graph in the validation set or test set; g·s Represents the label of the sth training graph in the training set before normalization, SOH g·s is the normalized label of the sth training graph in the training set, SOH g·j Represents the label of the jth graph in the validation set or test set before normalization, SOH g·j is the normalized label of the jth graph in the validation set or test set. EOL is the standard for health status transition, which is mostly used for remaining life prediction and is set between 0.7 and 0.8.

[0025] Furthermore, in step S4, a graph attention mechanism or a graph sample and aggre graph attention neural network (GAT) (GraphSAGE) or a graph convolutional neural network (GCN) is used to train the training set G using gradient descent. trainPerform training and learning to obtain a trained graph attention model or graph sample and aggre graph attention neural network (GAT) (GraphSAGE) model or graph convolutional neural network (GCN) model.

[0026] Furthermore, the test set is fed into a trained graph attention model or a graph sample and aggre graph attention networks (GAT) (GraphSAGE) model or a graph convolutional neural network (GCN) model to obtain the health status of the two half-cycles. At this time, denormalization is required as follows:

[0027]

[0028] SOH j is the predicted value of the health status of the j-th test image in the test set, is the predicted value of the health status of the j-th test image in the test set after denormalization.

[0029] Furthermore, in step S2, a complete charge-discharge cycle is regarded as a node, the health status of the two half cycles included in the charge-discharge cycle is used as the node feature, and the charge health status and discharge health status of the complete charge-discharge cycle are used as the node label; the number of nodes N, i.e., the window size, is set to establish a graph;

[0030] The charge health status and discharge health status of a complete charge and discharge cycle are the ratios of the charge capacity and discharge capacity to the rated capacity, respectively.

[0031] Furthermore, in step S2, in addition to constant current charging time, constant voltage charging time, and constant current discharging time, the node features also use health features extracted by artificial feature engineering, including equal voltage drop discharge time, equal voltage difference charging time, or discharge voltage sample entropy and other health features obtained through feature engineering.

[0032] Furthermore, in step S2, the duration of the corresponding charging and discharging process (stage) or the indirect health factor is used as the feature of the node; the indirect health factor refers to the feature extracted to characterize the health status of the battery.

[0033] The present invention has the following advantages and effects compared to the prior art:

[0034] (1) The impact of historical charge and discharge cycles on the health status of current or future charge and discharge cycles can be comprehensively considered, and the impact of each charge and discharge stage can also be comprehensively considered to improve prediction accuracy.

[0035] (2) It avoids the complicated work of feature extraction. Of course, manually extracted features can also be used.

[0036] (3) For the first time, we attempted to apply graph neural networks to the health status assessment of lithium-ion batteries, providing new research ideas for the health status assessment of lithium-ion batteries. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] In order to more clearly illustrate the technical solutions of the embodiments of the present application or the prior art, the following briefly introduces the drawings required for use in the embodiments or the prior art descriptions.

[0038] Figure 1 is an implementation flow chart in an embodiment of the present invention;

[0039] Figure 2 A diagram constructed in an embodiment of the present invention;

[0040] Figure 3 This is a schematic diagram of the prediction effect in Example 1 of the present invention;

[0041] Figure 4 This is a schematic diagram of the prediction effect in Example 2 of the present invention;

[0042] Figure 5 This is a schematic diagram of the prediction effect in Example 3 of the present invention. DETAILED DESCRIPTION

[0043] The core of the present invention is to provide a lithium-ion battery health status prediction method based on graph neural network, so that the health status assessment of lithium-ion batteries can improve the accuracy of prediction by introducing historical charge and discharge cycles and comprehensively considering various stages of the charge and discharge cycle.

[0044] In order to enable those skilled in the art to better understand the technical solutions of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. Obviously, the embodiments described are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present invention.

[0045] Example 1:

[0046] A method for predicting the health status of lithium-ion batteries based on graph neural networks, such as Figure 1 As shown, the following steps are included:

[0047] S1. Obtain the health status of each half cycle of the lithium-ion battery;

[0048] The charging and discharging process includes a charging process or a discharging process; the charging and discharging stage includes a constant current charging stage, a constant voltage charging stage or a constant current discharging stage; the charging process includes a constant current charging stage and a constant voltage charging stage, and the discharging process includes a constant current discharging stage;

[0049] Completing one charging process and one discharging process is a complete charge and discharge cycle, while a half cycle means completing one charging process or one discharging process alone.

[0050] The health status is the ratio of the charging capacity or discharging capacity to the rated capacity. The rated capacity is the factory parameter of the battery. It should be noted that the discharge capacity is a negative value in the present invention.

[0051] In this embodiment, the Oxford public data set (including 8 batteries, Cell1 to Cell8) is used as an example to collect the health status of the charging and discharging processes during each cycle.

[0052] S2, construct the nodes, node features and labels of the graph, and set the window size;

[0053] Consider a complete charge-discharge cycle as a node, and the health status of the two half cycles in the charge-discharge cycle as the node feature. Set the window size N=4, and take the health status of the N*2+1th half cycle after each N*2 half cycle as the label of the graph, thereby establishing a graph g. Through sliding windows, M isomorphic graphs g are formed. m , m=1~M, graph set G=g1,g2,g3,...g M .

[0054] In this embodiment, Figure 2 As shown, there are 4 nodes in the graph, that is, 4 complete cycles, each of which has 2 feature SOHs. f and SOH a , is the health status of the two adjacent half cycles in a complete cycle; and the connection between each node and other nodes constitutes an edge, and each graph corresponds to a health state SOH g , that is, the graph label, thereby establishing the graph g, and forming M isomorphic graphs g through the sliding window m , m=1~M, graph set G=g1,g2,g3,...g M .

[0055] S3. Divide the graph data into training set, validation set, and test set, and normalize the graph data;

[0056] M isomorphic graphs g m Divide into training set Gtrain =g1 train , g2 train , g3 train ,...g p train , where the subscripts train and subscripts represent the training set, p represents the number of samples in the training set, and g p train Represents the training set G train The p-th training graph in the validation set G valid =g1 valid , g2 valid , g3 valid ,...g b valid , where the subscripts valid and valid represent the validation set, b represents the number of samples in the training set, and g b valid Represents the training set G valid The b-th training graph in the test set G test =g1 test , g2 test , g3 test ,...g q teat , where the subscripts test and g represent the number of samples in the test set. q test Denotes the test set G test The g-th test image in , where p+b+q=M.

[0057] For the training set G train The training graph g in p train , find all training graphs g p train The maximum and minimum values ​​of the health status of the two half cycles (distinguished by f and a) in all nodes;

[0058] For the training set G train All nodes in all training graphs are normalized to extreme values:

[0059]

[0060]

[0061]

[0062] For the validation set G valid All validation graphs and test set G test All nodes in all test graphs are normalized to their extreme values:

[0063]

[0064]

[0065]

[0066] Among them, SOH f·s·t 、SOH a·s·t They represent the health status of the two half cycles before and after normalization of the t-th node in the s-th training graph in the training set; SOH f·j·k 、SOH a·j·k They represent the constant current charging time, constant voltage charging time, and constant current discharging time of the j-th verification graph or the k-th node in the test graph before normalization; SOH f·s·t 、SOH a·s·t is the normalized constant current charging time, constant voltage charging time, and constant current discharging time of the tth node in the sth training graph in the training set; SOH f·j·k 、SOH a·j·k SOH is the health status of the two half-cycles before and after normalization of the k-th node in the j-th graph in the validation set or test set; g·s Represents the label of the sth training graph in the training set before normalization, SOH g·s is the normalized label of the sth training graph in the training set, SOH g·j Represents the label of the jth graph in the validation set or test set before normalization, SOH g·j is the normalized label of the jth graph in the validation set or test set. EOL is the standard for health status transition, which is mostly used for remaining life prediction and is set between 0.7 and 0.8.

[0067] In this embodiment, all isomorphic graphs of Cell1 to Cell2 are taken as a training set, all isomorphic graphs of Cell3 to Cell4 are taken as a validation set, and all isomorphic graphs of Cell5 to Cell8 are taken as a test set.

[0068] S4. Use the graph neural network method to train and learn the training set data, and test the test set samples to achieve the health status prediction of lithium-ion batteries;

[0069] The graph sample and aggre graph attention networks (GAT) (GraphSAGE) method is used, and with the help of graph pooling and gradient descent, the training set G train Perform training and learning, and verify the set G validVerify and obtain a trained graph attention model or graph sample and aggre graph attention neural network (GAT) (GraphSAGE) model or graph convolutional neural network (GCN) model;

[0070] The test set is fed into the trained graph sample and aggre graph attention networks (GAT) (GraphSAGE) model to obtain the health status of the two half-cycles. At this time, denormalization is also required, as follows:

[0071]

[0072]

[0073] SOH j·k·c and SOH j·k·d is the predicted value of the health status of the charging and discharging process of the kth node in the jth test graph in the test set before denormalization, and is the predicted value of the health status of the charging and discharging process of the kth node in the jth test graph in the test set after denormalization; the EOL is 0.7.

[0074] Take the Cell5 prediction effect as an example, see Figure 3 ,As can be seen from the figure, the predicted value is consistent with the true value.

[0075] Example 2:

[0076] This embodiment is different from the embodiment 1 in that, in step S4, a graph attention neural network (GAT) method is adopted;

[0077] Take the Cell6 prediction effect as an example, see Figure 4 ,As can be seen from the figure, the predicted value is consistent with the true value.

[0078] Example 3:

[0079] This embodiment is different from the embodiment 1 in that, in step S4, a graph convolutional neural network (GCN) method is adopted;

[0080] Take the Cell7 prediction effect as an example, see Figure 5,As can be seen from the figure, the predicted value is consistent with the true value.

[0081] Example 4:

[0082] This embodiment is different from Embodiment 1 in that, in step S2, the characteristics of the node may be constant current charging time, constant voltage charging time, and constant current discharging time (or indirect health factors extracted through feature engineering).

[0083] The above embodiments are only specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention are included in the scope of protection of the present invention.

Claims

1. A method for predicting the health status of lithium-ion batteries based on graph neural networks, characterized in that: The following steps are involved: S1. Obtain the health status of each half cycle of the lithium-ion battery; S2. Construct the nodes, node features, and labels of the graph, and set the window size. Consider a complete charge-discharge cycle as a node, the health status of two adjacent half-cycles as node features, and the health status of the half-cycle after several charge-discharge cycles as the graph label. Set the number of nodes N, i.e. the window size, to build the graph g, and form M isomorphic graphs g through sliding windows. m , m=1~M, graph set G={g1,g2,g3,...g M }; S3, divide the graph data into training set, validation set and test set, and normalize the graph data; m Divide into training set G train ={g1 train ,g2 train ,g3 train ,...g p train }, where the subscripts train and subscripts represent the training set, p represents the number of samples in the training set, and g p train Represents the training set G train The p-th training graph in the validation set G valid ={g1 valid ,g2 valid ,g3 valid ,...g b valid }, where the subscripts valid and valid represent the validation set, b represents the number of samples in the training set, and g b valid Represents the training set G valid The b-th training graph in the test set G test ={g1 test ,g2 test ,g3 test ,...g q test }, where the subscripts test and q represent the number of samples in the test set, and g q test Denotes the test set G test The qth test graph in, where p+b+q=M; for the training set G train The training graph g in p train , find all training graphs g p train The maximum and minimum values ​​of the health status of all nodes in the two half cycles are distinguished by f and a; For the training set G train All nodes in all training graphs are normalized to extreme values: For the validation set G valid All validation graphs and test set G test All nodes in all test graphs are normalized to their extreme values: Among them, SOH f·s·t 、SOH a·s·t They represent the health status of the two half cycles before and after normalization of the t-th node in the s-th training graph in the training set; SOH f·j·k 、SOH a·j·k They represent the constant current charging time, constant voltage charging time, and constant current discharging time of the kth node in the jth verification graph or test graph in the test set before normalization. is the normalized constant current charging time, constant voltage charging time, and constant current discharging time of the tth node in the sth training graph in the training set; SOH is the health status of the two half-cycles before and after normalization of the k-th node in the j-th graph in the validation set or test set; g·s Represents the label of the s-th training graph in the training set before normalization, is the normalized label of the sth training graph in the training set, SOH g·j Represents the label of the jth graph in the validation set or test set before normalization, is the normalized label of the jth graph in the validation set or test set. EOL is the standard for health status transition, which is mostly used for remaining life prediction and is set between 0.7 and 0.

8. S4. Use the graph neural network method to train and learn the training set data, and test the test set samples to achieve the health status prediction of lithium-ion batteries.

2. The method for predicting the health status of a lithium-ion battery based on a graph neural network according to claim 1, characterized in that: In step S1, the charging and discharging process includes a charging process or a discharging process; the charging and discharging stage includes a constant current charging stage, a constant voltage charging stage or a constant current discharging stage; the charging process includes a constant current charging stage and a constant voltage charging stage, and the discharging process includes a constant current discharging stage; Completing one charging process and one discharging process is a complete charge and discharge cycle, while a half cycle means completing one charging process or one discharging process alone. The health status is the ratio of the charging capacity or discharging capacity to the rated capacity, where the rated capacity is the factory parameter of the battery.

3. The method for predicting the health status of a lithium-ion battery based on a graph neural network according to claim 1, wherein: In step S4, a graph attention mechanism or a graph sample and aggre graph attention neural network (GAT) (GraphSAGE) or a graph convolutional neural network (GCN) is used to train the training set G using gradient descent. train Perform training and learning to obtain a trained graph attention model or graph sample and aggre graph attention neural network (graph attentionnetworks, GAT) (GraphSAGE) model or graph convolutional neural network (GCN) model.

4. The method for predicting the health status of a lithium-ion battery based on a graph neural network according to any one of claims 1 to 3, characterized in that: The test set is fed into a trained graph attention model or a graph sample and aggre graph attention networks (GAT) (GraphSAGE) model or a graph convolutional neural network (GCN) model to obtain the health status of the two half-cycles. At this time, denormalization is required as follows: is the predicted value of the health status of the j-th test image in the test set, is the predicted value of the health status of the j-th test image in the test set after denormalization.

5. The method for predicting the health status of a lithium-ion battery based on a graph neural network according to claim 3, characterized in that: In step S2, a complete charge-discharge cycle is regarded as a node, the health status of the two half cycles included in the charge-discharge cycle is used as the node feature, and the charge health status and discharge health status of the complete charge-discharge cycle are used as the node label; the number of nodes N, i.e., the window size, is set to build a graph; The charge health status and discharge health status of a complete charge and discharge cycle are the ratios of the charge capacity and discharge capacity to the rated capacity, respectively.

6. The method for predicting the health status of a lithium-ion battery based on a graph neural network according to claim 3, characterized in that: In step S2, in addition to constant current charging time, constant voltage charging time, and constant current discharging time, the node features also include health features extracted by artificial feature engineering, including constant voltage drop discharging time, constant voltage difference charging time, and discharge voltage sample entropy.

7. The method for predicting the health status of a lithium-ion battery based on a graph neural network according to claim 3, characterized in that: In step S2, the duration of the corresponding charging and discharging process or the indirect health factor is used as the feature of the node; Indirect health factors refer to the features extracted to characterize the health status of the battery.

Images