Method and system for predicting and expressing graph structure of dissolved gas in multi-view transformer oil

Abstract

The invention relates to a multi-view transformer oil dissolved gas graph structure prediction representation method and system, and the method comprises the steps: collecting a multi-element gas sequence dissolved in oil, employing a multi-head attention mechanism module to extract hidden features, and obtaining a plurality of groups of RELI feature matrixes; a RELI feature matrix is segmented, a multi-view-graph generator is used to construct a feature graph, and a graph convolutional network is further used to process the feature graph to obtain features of different subspace information. And then, updating the RELI feature matrix, inputting the RELI feature matrix into the multi-head attention mechanism module again, and capturing a complex relationship among the features. Finally, the processed feature matrix is output to generate a predicted gas sequence. According to the method, the prediction accuracy is improved by 17.41%, the computing resource demand is remarkably reduced, the memory usage amount and the operation time are reduced, the feature extraction capability is optimized, the overfitting risk is reduced, the model efficiency is enhanced, and the method is suitable for the fields of electric power prediction, weather forecast, traffic flow prediction and the like.

Description

Technical Field

[0001] The invention belongs to the technical field of analysis of dissolved gas in oil-immersed transformer oil, and in particular relates to a method and system for predicting and representing a structure of a multi-view graph of dissolved gas in transformer oil. Background Art

[0002] Oil-immersed transformers are one of the important equipment in the power system, and their operating status directly affects the stability and reliability of the entire power system. In order to ensure the safe operation of the transformer, it is very necessary to regularly monitor and analyze the dissolved gas in the transformer oil. Dissolved Gas Analysis (DGA) is a commonly used diagnostic method that analyzes the various gas components and their concentrations dissolved in the transformer oil to assess whether there are abnormal conditions such as overheating and arc discharge inside the transformer.

[0003] To solve the problem of gas parameter sequence prediction, statistical methods are mainly used, such as the three ratio method, to identify fault types. Time series analysis, such as the ARIMA model and exponential smoothing method, is used to predict the trend of future gas concentration changes. Machine learning algorithms, such as support vector machines (SVM), decision trees, random forests, neural networks, etc., provide powerful tools for building prediction models to grasp the trend of gas concentration changes. At the same time, artificial intelligence technologies such as deep learning (such as convolutional neural networks CNN, recurrent neural networks RNN, etc.) are used to further improve prediction accuracy, opening up a new path for dissolved gas analysis in oil.

[0004] The fault diagnosis and early warning mechanism combining the expert system and the rule engine has pushed the monitoring of dissolved gas in oil-immersed transformer oil to a new level of greater accuracy and efficiency, helping to detect potential faults in a timely manner and take corresponding preventive measures to ensure the safe and stable operation of the power system.

[0005] However, facing the complex task of multi-gas variable sequence prediction, that is, predicting the future long-term sequence with the minimum error based on historical time series data, there are still many challenges. There are currently several problems with the multi-gas variable sequence prediction task, including the lack of relationship structure between multiple variables, weak ability to capture local dynamic changes in the entire long-term time series, high computational complexity, and unsatisfactory accuracy. In order to solve these problems, the framework of the present invention integrates GCN to process dynamically constructed graphs with topological information from three different perspectives: time view, dimension view, and cross view. Inspired by multi-head attention, the present invention designs a multi-head mechanism to improve the performance of the model to capture different features. Summary of the invention

[0006] In view of the above problems, the present invention proposes a method and system for predicting the structure of dissolved gas graphs in transformer oil using multiple views. The key point is the use of GCN, i.e., graph convolutional network, to extract the spatiotemporal structure and complex relationships between multi-element gases. GCN can aggregate information from adjacent parameters and filter out interference in the graph. Therefore, GCN has advantages in processing graph structure data. By constructing a graph structure learning model to dynamically generate graphs from parameter sequence data, GCN is used to optimize the information features of the processed data and the generated graph structure. The advantages of GCN are used to extract the features of multiple gas variables. By constructing a topological graph between multi-element variables, the features between the multi-element variables extracted from the graph can be complementary, thereby enhancing the model's ability to predict multi-element long-term sequences.

[0007] Existing models lack the ability to extract dependencies between multiple variables from multivariate long-term time series. Most of the existing models focus on extracting features in the time and frequency domains of the time series. Few of these models consider the relationship between multiple variables from a dimensional perspective. However, in most real datasets, there are meaningful relationships between different variables. At the same time, extracting these relationships is an important part of studying multivariate long-term time series prediction.

[0008] In a first aspect, the present invention provides a method for predicting and representing a structure of a dissolved gas graph in a multi-view transformer oil, which is characterized in that the method comprises the following steps:

[0009] Step (1) collecting multi-gas parameters dissolved in oil to construct a multi-gas sequence, wherein the multi-gas parameters include H2, CH4, C2H2, C2H4, and C2H6 gas parameters;

[0010] Step (2) extracting hidden features of the multi-gas sequence using a multi-head attention mechanism module to obtain multiple sets of RELI feature matrices; the multi-head attention mechanism module is composed of two sub-modules, RevIN and Linear-Group, wherein the RevIN sub-module maps the multi-gas input sequence to the multi-gas intermediate sequence through normalization and learnable parameters, and the Linear-Group sub-module maps the multi-gas intermediate sequence to multiple sets of RELI feature matrices;

[0011] Step (3) splitting the single RELI feature matrix output by the multi-head attention mechanism module into multiple sub-matrices one by one, and constructing multiple groups of feature graphs of each sub-matrix one by one using a multi-view-graph generator; the multi-view-graph generator includes a cross-graph generator, a dimensional graph generator and a time graph generator;

[0012] Step (4) uses GCN to process the feature graphs generated by the multi-view generator to obtain features containing different subspace information from the input sequence; for each graph generator, GCN is used to calculate the corresponding update factor and update the RELI feature matrix;

[0013] Step (5) inputs the updated multiple sets of RELI feature matrices into the multi-head attention mechanism module again for parallel processing to capture the complex relationship between the features and output one or more processed feature matrices; and uses the feature matrix to generate the final predicted gas sequence and optimize the prediction accuracy by minimizing the L2 loss function.

[0014] Furthermore, in step (1), the multi-gas sequence mainly includes gas parameter sequences such as H2, CH2, CH4, C2H2, C2H4, C2H6, and is defined as L x is the time step, d in is the number of gas categories.

[0015] Furthermore, in the step (2), the multi-head attention mechanism module specifically includes the following contents.

[0016] It mainly includes two sub-modules: RevIN module and parallel linear processing layer module (Linear-Group).

[0017] The RevIN module converts the multi-gas input sequence into Mapped to a multi-gas intermediate sequence X that conforms to the same distribution n .

[0018]

[0019] in, is the normalized multi-gas sequence, x i and Respectively represent the single gas sequence before and after normalization, is the mean of a single gas sequence, is the single gas sequence variance, λ1 and λ2 are two learnable parameters of the RevIN module.

[0020] The Linear-Group module converts the multi-gas intermediate sequence X into n Mapped into multiple sets of RELI feature matrices.

[0021]

[0022] in is the i-th RELI feature matrix, represents the i-th learnable parameter matrix, d is the mapping dimension, b i is the corresponding learnable bias.

[0023] Furthermore, in the step (3), a single RELI feature matrix is ​​divided into multiple sub-matrices, and the specific contents are as follows.

[0024] H is calculated by the following formula i Decompose into multiple sub-matrices of the same size, and adjacent sub-matrices have overlapping parts.

[0025]

[0026] in, l is the segmentation step length, represents the j-th sub-matrix, and Split(*) is the partitioning function.

[0027] Furthermore, in the step (3), the multi-view-graph generator and the multiple sets of feature graphs are specifically as follows.

[0028] The multi-view-graph generator includes a cross-graph generator, a dimensional graph generator, and a time graph generator.

[0029] Multiple groups of feature maps include cross feature maps, dimensional feature maps, and time feature maps.

[0030] (1) The input of the cross graph generator is two adjacent sub-matrices. The j-th sub-matrix is ​​obtained by the following formula: The corresponding first and second hidden layer sub-matrices and

[0031]

[0032] in, are two learnable weight matrices, b c1 and b c2 are two learnable bias terms, and Norm(*) represents the normalization function.

[0033] The first and second hidden layer sub-matrices are calculated by the following formula and The cross-feature map

[0034]

[0035] in Softmax2d is the relevant activation function.

[0036] (2) The input of the dimensional graph generator is a single submatrix. The j-th submatrix is ​​obtained by the following formula: The corresponding third hidden layer sub-matrix F n .

[0037]

[0038] in, is the difference matrix of the j-th submatrix, Diff(*) is the difference operation, and b d are the learnable matrix and bias term respectively, and Norm(*) represents the normalization function.

[0039] The dimensional feature map is obtained by the following formula

[0040]

[0041] in, for The transposed matrix of .

[0042] (3) The input of the time graph generator is a single submatrix. The j-th submatrix is ​​obtained by the following formula: Time characteristic diagram of

[0043]

[0044] in, is the jth sub-matrix The corresponding fourth hidden layer sub-matrix, for Transpose a matrix.

[0045] Furthermore, the updating of the RELI feature matrix in step (4) is specifically as follows.

[0046] Update the RELI feature matrix by the following formula.

[0047]

[0048] in, is the jth sub-matrix The mapping submatrix of is the jth sub-matrix The update factor of the time graph generator, is a learnable parameter and Concat is a connection function. is the updated i-th RELI feature matrix, and h is the number of RELI feature matrices.

[0049] Furthermore, the specific content of step (5) is as follows.

[0050] The final predicted gas sequence is obtained by the following formula

[0051]

[0052] in, represents the i-th learnable parameter matrix, d is the mapping dimension, b i is the corresponding learnable bias, is the mean of a single gas sequence, is the single gas sequence variance, λ1 and λ2 are two learnable parameters of the RevIN module, Multivariate gas prediction sequence before normalization, For the prediction results.

[0053] In a second aspect, the present invention provides a system for predicting and representing a structure of a dissolved gas graph in a multi-view transformer oil, which is characterized in that it includes:

[0054] The multi-head attention mechanism module is configured as follows:

[0055] Receiving a multi-gas sequence as input; applying a multi-head attention mechanism to extract hidden features of the multi-gas sequence, thereby generating multiple sets of initial RELI feature matrices; updating the multiple sets of initial RELI feature matrices and inputting them again into the multi-head attention mechanism module to further process and generate a final prediction sequence;

[0056] Multi-view processing module, configured as:

[0057] Receiving a single RELI feature matrix from the multi-head attention mechanism module as input; dividing the single RELI feature matrix into multiple sub-matrices one by one, wherein each sub-matrix represents a different view or dimension; and constructing multiple groups of feature maps of a single sub-matrix using a multi-view-graph generator for each sub-matrix one by one, wherein each group of feature maps corresponds to a specific view of the sub-matrix; processing each group of feature maps generated by the multi-view-graph generator using GCN to obtain an update factor; and updating the RELI feature matrix one by one using the update factor to obtain a RELI feature matrix processed by multiple views.

[0058] Compared with the prior art, the present invention has the following beneficial effects:

[0059] Improved prediction accuracy: The MT-GSR method has improved prediction accuracy by 17.41% compared to existing technologies. This means that more accurate prediction results can be provided in areas such as power forecasting, weather forecasting, and traffic flow forecasting. It reduces computing resource requirements and can significantly reduce memory usage by 66.52%, while reducing running time by up to 78.09%. It also performs well in computational efficiency, which helps to reduce hardware costs and improve the response speed of the system. It optimizes feature extraction capabilities. By using a dimensional graph generator based on a graph convolutional network (GCNS) to dynamically learn the structural relationship between variables in multivariate long-term time series, it can extract more effective information features, further improve the accuracy of predictions, reduce redundant information, thereby reducing the risk of overfitting, and achieve linear complexity, further enhancing the efficiency of the model. It adopts a multi-head mechanism, inspired by the multi-head self-attention mechanism in the Transformer, and uses a multi-head mechanism to optimize the features of the input sequence, thereby improving prediction performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Figure 1 The work flow chart of the method of the present invention is

[0061] Figure 2 The system framework diagram of the present invention is

[0062] Figure 3 The feature map-heat map visualization of the present invention

[0063] Figure 4 Comparison of the running memory of the method of the present invention and various methods DETAILED DESCRIPTION

[0064] The method and system of the present invention are further described below in conjunction with the accompanying drawings and embodiments, but this should not limit the scope of protection of the present invention.

[0065] This embodiment discloses a method and system for predicting the representation of dissolved gas graph structure in multi-view transformer oil, which utilizes the advantages of graph convolutional networks (GCN) to extract the spatiotemporal structure and complex relationships between multiple gas variables, thereby enhancing the model's ability to predict multivariate long-term series.

[0066] like Figure 1 As shown, a method for predicting and representing a structure of a dissolved gas graph in a multi-view transformer oil specifically comprises:

[0067] Step (1) uses a multi-parameter coupled oil-immersed transformer intelligent sensor device based on a micro-electromechanical system to collect a multi-gas sequence dissolved in the oil, including H2, CH2, CH4, C2H2, C2H4, C2H6, etc.

[0068] These gas sequences are defined as L x is the time step, d in is the number of gas categories.

[0069] Step (2) uses a multi-head attention mechanism module to extract the hidden features of the multi-gas sequence and obtain multiple sets of RELI feature matrices;

[0070] The multi-head attention mechanism module consists of two sub-modules, RevIN and Linear-Group. The RevIN sub-module maps the multi-gas input sequence to the multi-gas intermediate sequence through normalization and learnable parameters, and the Linear-Group sub-module maps the multi-gas intermediate sequence to multiple sets of RELI feature matrices, aiming to provide stable and information-rich functions for the subsequent graph generator. The detailed structure is shown in Figure 2 shown.

[0071] The RevIN submodule inputs the multi-gas into the sequence Mapped to a multi-gas intermediate sequence X that conforms to the same distribution n , the formula is as follows:

[0072]

[0073] In the formula, is the normalized multi-gas sequence, x i and Respectively represent the single gas sequence before and after normalization, is the mean of a single gas sequence, is the single gas sequence variance, λ1 and λ2 are two learnable parameters.

[0074] The Linear-Group submodule converts the multi-gas intermediate sequence X n Mapped into multiple sets of RELI feature matrices, the formula is as follows:

[0075]

[0076] In the formula, is the i-th RELI feature matrix, represents the i-th learnable parameter matrix, d is the mapping dimension, b i is the corresponding learnable bias.

[0077] Step (3) divides the single RELI feature matrix output by the multi-head attention mechanism module into multiple sub-matrices one by one, and constructs multiple sets of feature maps of a single sub-matrix using the multi-view-graph generator one by one, such as Figure 2 The purpose of this split is to reduce the model size, avoid overfitting, and allow features to be extracted from multiple perspectives.

[0078] The specific segmentation method is as follows:

[0079] Using a fixed segmentation step size l, the RELI feature matrix H is converted into i Dividing into multiple sub-matrices of the same size, with a certain overlap between adjacent sub-matrices, helps capture the local features in the time series and the continuity between them:

[0080]

[0081] In the formula, l is the segmentation step length, represents the jth submatrix, Split(*) is the partitioning function that represents the partitioning process, and the output is a series of submatrices {H j_1 ,H j_2 ,...,H j_n}, where j represents the index of the submatrix.

[0082] The multi-view-graph generator includes a cross-graph generator, a dimension-graph generator and a time-graph generator, which extract the features of the multi-element gas time series from different angles to form multiple sets of feature graphs.

[0083] ① Cross-graph generator: The input is two adjacent sub-matrices, which are passed through two hidden layers (the j-th sub-matrix The corresponding first and second hidden layer sub-matrices and ), calculate the cross feature map The formula is as follows:

[0084]

[0085] In the formula, are two learnable weight matrices, b c1 and b c2 are two learnable bias terms, and Norm(*) represents the normalization function.

[0086] Calculate the first and second hidden layer sub-matrices and The cross-feature map The formula is as follows:

[0087]

[0088] In the formula, Softmax2d is the relevant activation function.

[0089] ② Dimensional graph generator: The input is a single submatrix, the difference matrix is ​​calculated, and a hidden layer (the jth submatrix) is passed through it. The corresponding third hidden layer sub-matrix Fn ) Calculate the dimension feature map The formula is as follows:

[0090]

[0091] In the formula, is the difference matrix of the j-th submatrix, Diff(*) is the difference operation, and b d are the learnable matrix and bias term respectively, Norm(*) represents the normalization function;

[0092]

[0093] In the formula, for The transposed matrix of .

[0094] ③Time map generator: The input is a single sub-matrix, which is passed through a hidden layer (the jth sub-matrix The corresponding fourth hidden layer sub-matrix ) Calculate the j-th sub-matrix Time characteristic diagram of The formula is as follows:

[0095]

[0096] In the formula, is the jth sub-matrix The corresponding fourth hidden layer sub-matrix, for Transpose a matrix.

[0097] Step (4) uses a graph convolutional network (GCN) to process the feature map produced by the multi-view generator to obtain features containing different subspace information from the input sequence, such as Figure 2 As shown:

[0098] Step 4.1) For each graph generator (cross graph, dimension graph, time graph), use GCN to calculate the corresponding update factor.

[0099] Calculate the update factor of the submatrix corresponding to the cross graph generator The formula is as follows:

[0100]

[0101] In the formula, is the first hidden layer update factor, for The normalized adjacency matrix of for The degree matrix of for The element in the i-th row and j-th column of for The i-th element of the diagonal, is a learnable matrix, ReLU is the activation function, and b nlc is a learnable parameter;

[0102] Calculate the update factor of the submatrix corresponding to the dimensional graph generator The formula is as follows:

[0103]

[0104] In the formula, for The normalized adjacency matrix of for The degree matrix of and b nld is a learnable parameter, is the update factor of the second hidden layer, is the update factor of the submatrix corresponding to the cross graph generator.

[0105] Compute the update factor of the submatrix corresponding to the time graph generator The formula is as follows:

[0106]

[0107] In the formula, The third hidden layer update factor, for The normalized adjacency matrix of for The degree matrix of b o and b nlt is a learnable parameter, It is the update factor of the sub-matrix corresponding to the dimensional graph generator.

[0108] Step 4.2) Update the RELI feature matrix, the formula is as follows:

[0109]

[0110] In the formula, is the jth sub-matrix The mapping submatrix of is the jth sub-matrix The update factor of the time graph generator, is a learnable parameter, Concat is a connection function, is the updated i-th RELI feature matrix, and h is the number of RELI feature matrices.

[0111] Step (5) inputs the updated multiple sets of RELI feature matrices into the multi-head attention mechanism module again, and processes these feature matrices in parallel to capture the complex relationship between the features. Output one or more processed feature matrices, which contain information for generating the final predicted gas sequence, that is, the RevIN module uses the learned parameters to normalize and adjust the output of the multi-head attention mechanism module to generate the final prediction sequence, such as Figure 2 As shown:

[0112] Calculate the final predicted gas sequence, the formula is as follows:

[0113]

[0114] In the formula, represents the i-th learnable parameter matrix, d is the mapping dimension, b i is the corresponding learnable bias, is the mean of a single gas sequence, is the single gas sequence variance, λ1 and λ2 are two learnable parameters of the RevIN module, which are used to adjust the strength and direction of normalization. Multivariate gas prediction sequence before normalization, is the prediction result, i.e. the multi-gas sequence adjusted by the RevIN module.

[0115] The loss function L is the L2 loss value between the final prediction P and the true value Y, and the formula is as follows:

[0116]

[0117] In the formula, p i is the i-th predicted value in the predicted result P, y i is the ith value of the actual result. By minimizing the L2 loss function, its parameters can be gradually adjusted to more accurately predict the gas sequence.

[0118] Application examples:

[0119] In order to evaluate the performance of the model of the present invention, experiments were conducted on two available datasets A and B for 220V oil-immersed transformers. A sliding window is used to select each set of data on the dataset. The step size between each set of data is 1. The size of the sliding window is the sum of the input length Lx and the prediction length Lp. Set Lx=96, Lp∈{96, 192, 336, 720}. The mean square error (MSE) and mean absolute error (MAE) are selected as evaluation indicators. The learning rate is initialized to 0.001, and the batch size is set to 64. At the same time, 2 of the latest transformer-based methods and 1 of the latest GNN-based methods are selected as comparative experiments.

[0120] Reformer extends the transformer with LSH attention based on the hashing algorithm;

[0121] Log-Trans introduces a sparse attention called Log-Sparse Attention, which selects time steps that follow an exponential growth interval;

[0122] MTGNN uses a novel extended inception layer and hybrid jump propagation layer to refine the spatiotemporal features in gas sequences.

[0123] 1.1 Evaluation Results

[0124] The length of the gas sequence Lx is set to a constant, and the effectiveness of the proposed method and all the comparison methods on four prediction lengths Lp are evaluated: 96, 192, 336, and 720. The results of the main experiments are shown in Table 1. In all cases, the proposed method produces better results than other models or methods. Compared with the Reformer-based MTGNN, the improvement of the proposed method is particularly obvious on datasets with strong correlations between multi-gas variables, such as dataset A, which are 31.0%, 42.5%, 43.6%, and 55.2%, respectively.

[0125] Table 1 Evaluation comparison results of multiple methods

[0126]

[0127]

[0128] 1.2 Hyperparameter Evaluation

[0129] We conducted extended experiments to study the hyperparameters of the described inventive method, including the sub-matrix size of the multi-view-graph generator and the number of linear processing layer modules of the multi-head attention mechanism module.

[0130] In terms of submatrix size, we selected three values ​​12, 24, and 48 and conducted experiments using dataset A, as shown in Table 2. For dataset A, the best performing hyperparameter is set to 24.

[0131] Table 2 Prediction comparison results of different submatrix sizes

[0132]

[0133] Table 3 Prediction comparison results of the number of linear processing layer modules of different multi-head attention mechanism modules

[0134]

[0135] On the other hand, regarding the number of linear processing layer modules of the multi-head attention mechanism module, Table 3 shows that when the number is 8, the performance of the inventive method achieves the best effect and as shown in the accompanying figure,

[0136] 1.3Heat-map

[0137] In the attached Figure 3 In sub-figures (a) and (d) in Figure 2, the gas parameters of dataset A have richer periodic variables, so it has a more dispersed distribution of time series characteristics compared to the dataset B with rich gas trend components. Figure 3 In sub-figures (b) and (e) in Figure 2, the dimension plots show different relationships for each gas variable. Unlike the other two feature plots, the cross-feature plots are asymmetric because they calculate the mutual correlation between the two sub-matrices. Figure 3 (c) and (f) show the fluctuation relationship between adjacent sub-matrices with different local trends and periodic changes.

[0138] 1.4 Computational Efficiency

[0139] Attached Figure 4 It shows that the invented method is a linear model with a greatly reduced model size compared to Reformer and other duali models. With a prediction length of 1800, the memory usage of the invented method is 66.52% less than that of Reformer.

Claims

1. A method for predicting and representing a structure of a multi-view graph of dissolved gas in transformer oil, characterized in that: This includes the following steps: Step (1) collecting multi-gas parameters dissolved in oil to construct a multi-gas sequence, wherein the multi-gas parameters include H2, CH4, C2H2, C2H4, and C2H6 gas parameters; Step (2) extracting hidden features of the multi-gas sequence using a multi-head attention mechanism module to obtain multiple sets of RELI feature matrices; the multi-head attention mechanism module is composed of two sub-modules, RevIN and Linear-Group, wherein the RevIN sub-module maps the multi-gas input sequence to the multi-gas intermediate sequence through normalization and learnable parameters, and the Linear-Group sub-module maps the multi-gas intermediate sequence to multiple sets of RELI feature matrices; Step (3) splitting the single RELI feature matrix output by the multi-head attention mechanism module into multiple sub-matrices one by one, and constructing multiple groups of feature graphs of each sub-matrix one by one using a multi-view-graph generator; the multi-view-graph generator includes a cross-graph generator, a dimensional graph generator and a time graph generator; Step (4) uses GCN to process the feature graphs generated by the multi-view generator to obtain features containing different subspace information from the input sequence; for each graph generator, GCN is used to calculate the corresponding update factor and update the RELI feature matrix; Step (5) inputs the updated multiple sets of RELI feature matrices into the multi-head attention mechanism module again for parallel processing to capture the complex relationship between the features and output one or more processed feature matrices; and uses the feature matrix to generate the final predicted gas sequence and optimize the prediction accuracy by minimizing the L2 loss function.

2. The method for predicting and representing the structure of a multi-view graph of dissolved gas in transformer oil according to claim 1, characterized in that: In step (1), the multi-gas sequence is defined as L x is the time step, d in is the number of gas categories.

3. The method for predicting and representing the structure of a multi-view graph of dissolved gas in transformer oil according to claim 1, characterized in that: In step (2), the RevIN module inputs the multi-gas into the sequence Mapped to a multi-gas intermediate sequence X that conforms to the same distribution n , the formula is as follows: x n =λ1×X m +λ2 In the formula, is the normalized multi-gas sequence, x i and Respectively represent the single gas sequence before and after normalization, is the mean of a single gas sequence, is the single gas sequence variance, λ1 and λ2 are two learnable parameters.

4. The method for predicting and representing the structure of a multi-view graph of dissolved gas in transformer oil according to claim 1, characterized in that: In step (2), the Linear-Group module converts the multi-gas intermediate sequence X n Mapped into multiple sets of RELI feature matrices, the formula is as follows: In the formula, is the i-th RELI feature matrix, represents the i-th learnable parameter matrix, d is the mapping dimension, b i is the corresponding learnable bias.

5. The method for predicting and representing the structure of dissolved gas graphs in multi-view transformer oil according to claim 1, characterized in that: In step (3), each RELI feature matrix H is converted into i Split into multiple sub-matrices of the same size, with adjacent sub-matrices overlapping: In the formula, l is the segmentation step length, S j represents the j-th sub-matrix, and Split(*) is the partitioning function.

6. The method for predicting and representing the structure of dissolved gas graphs in multi-view transformer oil according to claim 1, characterized in that: The cross map generator calculates the cross feature map through two hidden layer sub-matrices The formula is as follows: In the formula, and is the jth sub-matrix The corresponding first and second hidden layer sub-matrices, are two learnable weight matrices, b c1 and b c2 are two learnable bias terms, Norm(*) represents the normalization function; Softmax2d is the relevant activation function.

7. The method for predicting and representing the structure of a multi-view graph of dissolved gas in transformer oil according to claim 1, characterized in that: The dimension map generator is constructed by using the difference matrix F d and the third hidden layer sub-matrix F n， Calculate dimensional feature map The formula is as follows: In the formula, is the difference matrix of the j-th submatrix, Diff(*) is the difference operation, and b d are the learnable matrix and bias term, respectively, n is the jth sub-matrix The corresponding third hidden layer sub-matrix, Norm(*) represents the normalization function; is the third hidden layer sub-matrix The transposed matrix of .

8. The method for predicting and representing the structure of a multi-view graph of dissolved gas in transformer oil according to claim 1, characterized in that: The time map generator is obtained by the fourth hidden layer sub-matrix Computation time characteristic graph The formula is as follows: In the formula, is the jth sub-matrix The corresponding fourth hidden layer sub-matrix is, is the fourth hidden layer sub-matrix Transpose a matrix.

9. The method for predicting and representing the structure of a multi-view graph of dissolved gas in transformer oil according to claim 1, characterized in that: The step (4) uses GCN to calculate the corresponding update factor, which specifically includes: Calculate the update factor of the submatrix corresponding to the cross graph generator The formula is as follows: In the formula, is the first hidden layer update factor, for The normalized adjacency matrix of for The degree matrix of for The element in the i-th row and j-th column of for The i-th element of the diagonal, is a learnable matrix, ReLU is the activation function, and b nlc is a learnable parameter; Calculate the update factor of the submatrix corresponding to the dimensional graph generator The formula is as follows: In the formula, for The normalized adjacency matrix of for The degree matrix of and b nld is a learnable parameter, is the update factor of the second hidden layer, is the update factor of the submatrix corresponding to the cross graph generator; Compute the update factor of the submatrix corresponding to the time graph generator The formula is as follows: In the formula, The third hidden layer update factor, for The normalized adjacency matrix of for The degree matrix of b o and b nlt is a learnable parameter, It is the update factor of the sub-matrix corresponding to the dimensional graph generator.

10. The method for predicting and representing the structure of dissolved gas graphs in multi-view transformer oil according to claim 1, characterized in that: In step (4), the RELI feature matrix is ​​updated according to the following formula: In the formula, S j ′ is the jth sub-matrix The mapping submatrix of is the jth sub-matrix The update factor of the time graph generator, is a learnable parameter, Concat is a connection function, is the updated i-th RELI feature matrix, and h is the number of RELI feature matrices.

11. The method for predicting and representing the structure of a multi-view graph of dissolved gas in transformer oil according to claim 1, characterized in that: The step (5) is to calculate the final predicted gas sequence, and the formula is as follows: In the formula, represents the i-th learnable parameter matrix, d is the mapping dimension, b i is the corresponding learnable bias, is the mean of a single gas sequence, is the single gas sequence variance, λ1 and λ2 are two learnable parameters of the RevIN module, Multivariate gas prediction sequence before normalization, For the prediction results.

12. A system for predicting and representing the structure of a multi-view graph of dissolved gas in transformer oil, characterized in that: include: The multi-head attention mechanism module is configured as follows: Receiving a multi-gas sequence as input; applying a multi-head attention mechanism to extract hidden features of the multi-gas sequence, thereby generating multiple sets of initial RELI feature matrices; updating the multiple sets of initial RELI feature matrices and inputting them again into the multi-head attention mechanism module to further process and generate a final prediction sequence; Multi-view processing module, configured as: Receiving a single RELI feature matrix from the multi-head attention mechanism module as input; splitting the single RELI feature matrix into multiple sub-matrices one by one, wherein each sub-matrix represents a different view or dimension; and constructing multiple sets of feature maps of a single sub-matrix using a multi-view-graph generator for each sub-matrix one by one, wherein each set of feature maps corresponds to a specific view of the sub-matrix; processing each set of feature maps generated by the multi-view-graph generator using a GCN to obtain an update factor; The RELI feature matrix is ​​updated one by one using the update factor, so as to obtain a RELI feature matrix after multi-view processing.

Images