A power transformer state evaluation method based on multi-metric residual graph convolution network
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- WUHAN UNIV
- Filing Date
- 2024-08-07
- Publication Date
- 2026-07-24
AI Technical Summary
Existing deep learning-based transformer state assessment methods suffer from insufficient data correlation mining when dealing with multivariate, nonlinear, and time-varying data, which affects the accuracy of assessment.
A multi-metric residual graph convolutional network is used to model the graph structure of dissolved gas data in oil of power transformers. A k-nearest neighbor graph is constructed through a multi-metric fusion graph construction method. A residual graph convolutional network is constructed using graph convolutional layers, batch processing layers, ReLU activation function and graph pooling layer to extract features and realize state assessment.
It improves the accuracy of transformer condition assessment, enables timely online condition monitoring, avoids equipment damage, and ensures the safe and stable operation of the power grid.
Smart Images
Figure CN118940120B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of power transformer condition monitoring and health assessment technology, and more specifically, relates to a method for assessing the health status of power transformers. Background Technology
[0002] Transformers, as core equipment in modern power grids, play a vital role in power transmission and voltage conversion. Their safe and stable operation is crucial to the reliability and stability of the power grid. Transformer operating conditions are complex; any failure can disrupt the normal operation of the power system and even cause serious economic losses. Therefore, monitoring the operating status of transformers is of great significance for ensuring the safety and stability of the power grid and reducing equipment damage.
[0003] Transformer condition monitoring methods can be mainly divided into offline preventive monitoring and online monitoring methods. Offline monitoring methods include electrical tests such as partial discharge, but these methods have a certain time lag. In recent years, deep learning-based methods have been widely used in the field of condition assessment due to their outstanding advantages in extracting deep feature representations. However, deep learning-based models must use regular Euclidean data as input, and when the transformer's operating condition is abnormal, multiple monitoring data interact and change simultaneously, exhibiting time-varying and nonlinear characteristics. These methods still have shortcomings in mining variable correlations.
[0004] As the operating environment of transformers changes, dissolved gases in the oil and other monitoring data can interact through decomposition and synthesis. Therefore, exploring the correlations among multivariate data is crucial for improving model performance. Graph neural networks, using non-Euclidean data as input, can better model the spatial relationships between variables. To this end, graph structure modeling is performed on power transformer operation monitoring data to fully explore the spatial correlations of the data. A condition monitoring scheme based on multi-metric residual graph convolution is proposed to improve the accuracy of condition assessment. Summary of the Invention
[0005] To address the aforementioned deficiencies or improvement needs of existing technologies, this invention proposes a power transformer condition assessment method based on a multi-metric residual graph convolutional network, thereby improving the accuracy of condition assessment.
[0006] To achieve the above objectives, this invention provides a power transformer state assessment method based on a multi-metric residual graph convolutional network, comprising:
[0007] Data on dissolved gases in oil under different health conditions of transformers were obtained, and the data were divided into training set and test set. Then, the data in the training set was preprocessed to balance.
[0008] The data in the training and test sets after balancing preprocessing are subjected to max-min normalization, and then a multi-metric fusion graph construction method is used to construct a k-nearest neighbor graph from the normalized training and test sets.
[0009] Feature extraction units are constructed based on graph convolutional layers, batch processing layers, ReLU activation function, and graph pooling layers. A residual graph convolutional network is then built based on multiple feature extraction units as a state evaluation model.
[0010] The k-nearest neighbor graphs of the training and test sets are input into the constructed residual graph convolutional network to train the state evaluation model, so as to evaluate the operating state of the transformer through the trained state evaluation model.
[0011] In some optional implementations, the process of acquiring dissolved gas data in the oil under different health conditions of the transformer, dividing the data into training and test sets, and then performing balanced preprocessing on the data in the training set includes:
[0012] The dissolved gas data in the oil of the transformer's historical operation data were obtained, namely H2, CH4, C2H6, C2H4, and C2H2. The dissolved gas data in the oil were divided into multiple operating states according to the size of the health index, and then divided into training set and test set according to the proportion.
[0013] The SMOTE Tomek mixed sampling algorithm is used to process the data in the training set to achieve a balance among the various types of data in the training set.
[0014] In some optional implementations, the step of performing max-min normalization on the pre-processed training and test sets, and then constructing a k-nearest neighbor graph from the normalized training and test sets using a multi-metric fusion graph construction method, includes:
[0015] The max-min normalization method is used to normalize the data in the training and test sets after the balanced preprocessing.
[0016] Each monitored variable is treated as a node, and the correlation between monitored variables is treated as an edge graph, where the monitored variable represents dissolved gas in oil;
[0017] The correlation between each node is calculated using the Gaussian kernel function, cosine similarity, and Pearson correlation coefficient. The average of the three adjacency matrices is then taken to fuse multiple metrics and obtain the fused adjacency matrix.
[0018] For the merged adjacency matrix, each node retains only the k nodes with the highest correlation to that node, which are used as the final correlation matrix A to construct the k nearest neighbor graph.
[0019] In some optional implementations, the construction of feature extraction units based on graph convolutional layers, batch processing layers, ReLU activation functions, and graph pooling layers, and the construction of a residual graph convolutional network as a state evaluation model based on multiple feature extraction units, includes:
[0020] Feature extraction units are constructed based on graph convolutional layers, batch processing layers, and ReLU activation functions, and each feature extraction unit is connected to an EdgePool graph pooling layer.
[0021] A feature extraction network is constructed by stacking several feature extraction units and adding a residual connection to every other feature extraction unit. The extracted features are input to a fully connected layer and classified using a softmax function to output the state evaluation result.
[0022] The constructed k-nearest neighbor graph is input into the residual graph convolutional network to obtain the state evaluation result, thus completing the training of the state evaluation model.
[0023] In some alternative implementations, the EdgePool graph pooling layer output is: X output =EdgePool(X′), where X′ = Relu(BN(GConv(AXW))), A is the adjacency matrix, X is the input data, and W is the trainable weight matrix.
[0024] In some optional implementations, evaluating the transformer's operating state using a trained state evaluation model includes:
[0025] After performing max-min normalization on the transformer operating data, a multi-metric fusion graph construction method is used to construct a k-nearest neighbor graph from the pre-balanced transformer operating data. The k-nearest neighbor graph data of the transformer operating data is then input into the trained state assessment model to obtain the online monitoring results of the transformer operating status.
[0026] In summary, compared with the prior art, the above-described technical solutions conceived by this invention can achieve the following beneficial effects:
[0027] This invention proposes a power transformer condition assessment method based on a multi-metric residual graph convolutional network. It models the dissolved gases in power transformer oil using a graph structure to fully exploit the spatial correlations of the data. A multi-metric fusion graph construction method is used to build a k-nearest neighbor graph of multiple monitoring variables, which is then input into the constructed residual graph convolutional network for feature extraction. This enables online condition assessment of the power transformer, allowing for timely processing and maintenance to prevent serious damage. Attached Figure Description
[0028] Figure 1 This is a schematic diagram of a method provided in an embodiment of the present invention;
[0029] Figure 2 This is a schematic diagram of a multi-metric fusion mapping method provided in an embodiment of the present invention;
[0030] Figure 3 This is a schematic diagram of a residual graph convolutional network provided in an embodiment of the present invention. Detailed Implementation
[0031] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.
[0032] like Figure 1 As shown, a power transformer state assessment method based on a multi-metric residual graph convolutional network is presented. This method comprises five parts: data balancing preprocessing, multi-metric fusion graph construction, spatiotemporal feature learning, offline model training, and online state assessment. Specifically, it includes the following steps:
[0033] Step S1: Obtain dissolved gas data in oil under different health conditions of the transformer, divide the data into training set and test set, and then perform balanced preprocessing on the data in the training set;
[0034] Furthermore, step S1 can be implemented in the following way:
[0035] Step S11: Obtain the dissolved gas data in the oil of the transformer's historical operating data, namely H2, CH4, C2H6, C2H4, and C2H2. According to the size of the health index, they are divided into five operating states: very good, good, moderate, poor, and very poor. They are divided into training set and test set according to a ratio of 70% and 30%, respectively.
[0036] For example, the dataset can be divided into five operating states: very good, good, moderate, poor, and very poor, based on the health index ranging from 0.8 to 1, 0.6 to 0.8, 0.4 to 0.6, 0.2 to 0.4, and 0 to 0.2.
[0037] Step S12: The training set data is processed using the SMOTE Tomek mixture sampling algorithm to balance the various types of data in the training set. The specific steps are as follows:
[0038] Step S121: Randomly select a subset of samples x from the minority class, calculate the Euclidean distance between it and other samples in the minority class sample set, and obtain its k nearest neighbors;
[0039] Step S122: Randomly select several samples x from the k nearest neighbors. nFor the randomly selected nearest neighbor sample x n According to x new =x + rand(0,1) * |xx n | Construct new samples and add them to the minority class samples;
[0040] Step S123: For most categories, calculate the nearest neighbor of each sample using Euclidean distance and determine whether the category labels of each pair of samples are the same;
[0041] Step S124: For samples of different categories, create a Tomek Link pair and mark it as noise, then remove it from the dataset.
[0042] Step S2: Perform max-min normalization on the data in the balanced preprocessed training and test sets, and then use the multi-metric fusion graph construction method to construct a k-nearest neighbor graph on the normalized training and test set data.
[0043] Furthermore, step S2 can be implemented in the following way:
[0044] Step S21: Normalize the divided training and test sets using the max-min normalization method. The calculation formula is as follows: Where x′ represents the value of a single data point, min is the minimum value in the column containing the data, and max is the maximum value in the column containing the data.
[0045] Step S22: Calculate the correlation between each node using Gaussian kernel function, cosine similarity and Pearson correlation coefficient respectively, and take the average of the three adjacency matrices. For each node, retain the k nodes with the highest correlation to that node as the final correlation matrix A to construct the k nearest neighbor graph.
[0046] Among them, multi-metric fusion graph construction methods such as Figure 2 As shown. For the edge weights between any two nodes, the similarity between nodes is calculated using the Gaussian kernel function, cosine similarity, and Pearson correlation coefficient, respectively. The Gaussian kernel function measures similarity by projecting vectors into a higher-order space, the cosine distance measures the difference between sample vectors, and the Pearson correlation coefficient measures the linear relationship between two variables. Node x i With node x j The formulas for calculating the Gaussian kernel function, cosine similarity, and Pearson correlation coefficient between the two pairs of strings are as follows:
[0047]
[0048] Where ξ is the bandwidth of the Gaussian kernel.
[0049]
[0050] The three calculated adjacency matrices are summed and their average is taken as the final correlation matrix. To reduce redundant connections, for each node, the k nodes with the largest weights are selected as its neighbors, and the edges connecting them are retained. Finally, redundant edges are deleted, i.e., their weights are set to 0, thus obtaining the correlation matrix A of the k-nearest neighbor graph.
[0051] Step S3: Construct feature extraction units based on graph convolutional layers, batch processing layers, ReLU activation function, and graph pooling layers, and build a residual graph convolutional network as a state evaluation model based on multiple feature extraction units;
[0052] like Figure 3 The schematic diagram of the residual graph convolutional network is shown below: The proposed residual graph convolutional network consists of 5 stacked feature extraction modules. Each feature extraction module consists of a graph convolutional network (GConv), a batch processing layer, and a ReLU activation function. The output of a single feature extraction unit can be expressed as X′=Relu(BN(GConv(AXW))), where A is the adjacency matrix, X is the input data, and W is the trainable weight matrix.
[0053] Each feature extraction unit is followed by an EdgePool graph pooling layer. The EdgePool graph pooling layer reduces the number of nodes, lowers computational requirements, and improves model performance. Therefore, the output data is input into the EdgePool layer for pooling, and its output can be represented as X. output =EdgePool(X′). Since deep graph convolutional networks tend to repeatedly apply Laplacian smoothing to different clusters, leading to a decrease in classification performance, residual connections are added every other feature extraction unit to prevent this. After feature extraction, the data is input to two fully connected layers and classified using a softmax classifier to obtain the final state evaluation result.
[0054] Step S4: Input the k-nearest neighbor graphs of the training set and test set data obtained in step S2 into the residual graph convolutional network constructed in step S3 to train the state evaluation model and evaluate the operating state of the transformer.
[0055] Step S5: For the online monitoring data of the transformer, the data is preprocessed in step S1, and then a k-nearest neighbor graph is constructed in step S2. The graph structure data of the online monitoring data is input into the trained state assessment model to realize the online state assessment of the power transformer.
[0056] It should be noted that, depending on the implementation needs, the various steps / components described in this application can be broken down into more steps / components, or two or more steps / components or parts of the operation of steps / components can be combined into new steps / components to achieve the purpose of this invention.
[0057] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
Claims
1. A method for assessing the state of a power transformer based on a multi-metric residual graph convolutional network, characterized in that, include: Data on dissolved gases in oil under different health conditions of transformers were obtained. The data were divided into training and test sets. The SMOTE Tomek hybrid sampling algorithm was used to process the data in the training set to balance the various types of data in the training set. The data in the balanced preprocessed training and test sets are subjected to min-max normalization. Then, a multi-metric fusion graph construction method is used to construct a k-nearest neighbor graph from the normalized training and test sets. Specifically, this includes: normalizing the data in the balanced preprocessed training and test sets using min-max normalization; treating each monitored variable as a node and the correlation between monitored variables as an edge graph, where the monitored variable represents dissolved gas in oil; calculating the correlation between each node using Gaussian kernel function, cosine similarity, and Pearson correlation coefficient, and averaging the three calculated adjacency matrices to fuse multiple metrics into a fused adjacency matrix; for the fused adjacency matrix, each node retains the k nodes with the highest correlation to that node as the final correlation matrix. To construct a k-nearest neighbor graph; A feature extraction unit is constructed based on graph convolutional layers, batch processing layers, ReLU activation function, and graph pooling layer. A residual graph convolutional network is then built based on these feature extraction units as a state evaluation model. Specifically, this includes: constructing feature extraction units based on graph convolutional layers, batch processing layers, and ReLU activation function, with each feature extraction unit connected to an EdgePool graph pooling layer; stacking several feature extraction units and adding residual connections between every other feature extraction unit to construct a feature extraction network; inputting the extracted features into a fully connected layer and classifying them using a softmax function to output the state evaluation result; and inputting the constructed k-nearest neighbor graph into the residual graph convolutional network to obtain the state evaluation result, thus completing the training of the state evaluation model. The k-nearest neighbor graphs of the training and test sets are input into the constructed residual graph convolutional network to train the state evaluation model, so as to evaluate the operating state of the transformer through the trained state evaluation model.
2. The method according to claim 1, characterized in that, The process involves acquiring dissolved gas data in the transformer oil under different health conditions, dividing the data into training and testing sets, including: The dissolved gas data in the oil of the transformer's historical operation data were obtained, namely H2, CH4, C2H6, C2H4, and C2H2. Based on the size of the health index, the dissolved gas data in the oil were divided into multiple operating states and then divided into training and test sets according to the proportion.
3. The method according to claim 2, characterized in that, The output of the EdgePool graph pooling layer is: ,in, , It is an adjacency matrix. For input data, is a trainable weight matrix.
4. The method according to claim 1, characterized in that, The process of evaluating the transformer's operating state using a trained state evaluation model includes: After performing max-min normalization on the transformer operating data, a multi-metric fusion graph construction method is used to construct a k-nearest neighbor graph from the pre-balanced transformer operating data. The k-nearest neighbor graph data of the transformer operating data is then input into the trained state assessment model to obtain the online monitoring results of the transformer operating status.