Transformer short-circuit current evaluation method based on adaptive multi-view neural network and related device

Abstract

The invention provides a transformer short-circuit current evaluation method based on an adaptive multi-view neural network and a related device, and the method comprises the steps: carrying out the load flow calculation through simulating different operation modes and state fluctuations, and obtaining original data as an input feature matrix; performing Z-score standardization processing on the feature matrix; dividing the standardized data into a training set and a test set; constructing a topological graph and a feature graph; constructing a power grid operation model based on the adaptive multi-view image convolutional neural network; training a power grid operation model by using the training set; and after completing model training, operating the power grid operation model on the test set to predict the short-circuit current of the transformer, and calculating an evaluation index to measure the model performance so as to evaluate the accuracy of predicting the short-circuit current by the model. According to the invention, the accuracy and response speed of short-circuit current prediction can be effectively improved, the safety and stability of a power grid are enhanced, and economic loss and maintenance cost caused by short-circuit faults are reduced at the same time.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence algorithms, and in particular to a transformer short-circuit current evaluation method and related devices based on an adaptive multi-view graph neural network. Background Art

[0002] Short-circuit faults are one of the most common types of failure in power systems. They not only damage equipment but can also trigger power outages, posing a serious threat to the safe and stable operation of the power grid. Transformers, as critical components in power systems, are subject to significant electrodynamic and thermal stresses when a short circuit occurs. Therefore, accurately assessing transformer short-circuit current is crucial for preventing transformer failures and protecting equipment. Accurately calculating short-circuit current allows for optimal selection of circuit breaker capacity and optimized relay protection configuration, effectively improving the reliability and cost-effectiveness of power systems.

[0003] Traditional assessment methods rely primarily on node characteristics, such as impedance and voltage levels, to calculate short-circuit currents. However, these methods often overlook the topological structure of the power grid, namely the connection relationships between components in the grid and their impact on the distribution of short-circuit currents. In complex power grid structures, this simplification can lead to significant deviations between the assessment results and actual conditions. Furthermore, traditional methods typically assume a static model of the power grid, which cannot adapt to the dynamic changes of modern power grids. The limitations of this static model are particularly evident in the context of large-scale renewable energy integration and the development of smart grids.

[0004] To address these challenges, some researchers have begun exploring the application of graph convolutional neural networks (GCNs) to short-circuit current assessment in power systems. Graph convolutional neural networks (GCNs) are a deep learning method based on graph data structures that can effectively process data with complex connectivity. In a power system, the power grid can be modeled as a graph, where nodes represent devices (such as generators, transformers, and loads) and edges represent the connections between devices. Compared to traditional methods, GCNs can simultaneously consider both node and topological features, thereby more accurately modeling the operational characteristics of the power system. The basic principle of GCNs is to update node feature representations by aggregating information about a node's local neighborhood, thereby capturing relationships and patterns within the graph structure. In a GCN, the features of each node depend not only on itself but also on the features of its immediate neighbors. Specifically, GCNs perform neighborhood aggregation, performing a weighted sum or mean calculation on the features of a node and its neighbors, and transforming them using a learnable weight matrix to generate a new node feature representation. This process is typically repeated across multiple layers of the network, with each layer gradually expanding the node's information receptive field, allowing the node to gradually acquire a wider range of graph structural information. In this way, GCN can effectively capture local and global structural features in graph data and is widely used in tasks such as node classification, link prediction, and graph embedding. In previous GCN-based short-circuit current assessment, although GCN has performed well in processing graph-structured data, it has the following shortcomings:

[0005] 1. GCN has limited ability to integrate node features and topology, which limits its performance in some classification tasks. When evaluating short-circuit current, this may lead to failure to fully consider the topological structure information in the power grid, thus affecting the accuracy of the evaluation.

[0006] 2. GCN may not be able to adaptively learn the deep correlation information between node features and topological structures, which means that it may not be able to adjust the weights according to the specific situation to better extract the most relevant information.

[0007] 3. GCN usually only considers the traditional topology and ignores the potential value of feature graphs. In real power grids, feature graphs (such as the k-nearest neighbor graph generated by node features) may reveal the intrinsic connections of the power grid better than traditional topology, which is particularly important for accurately assessing short-circuit current. Summary of the Invention

[0008] Based on the above problems, the present invention proposes a transformer short-circuit current assessment method and related devices based on an adaptive multi-view graph neural network, which can effectively improve the accuracy and response speed of short-circuit current prediction, enhance the safety and stability of the power grid, and reduce the economic losses and maintenance costs caused by short-circuit faults.

[0009] A transformer short-circuit current assessment method based on an adaptive multi-view graph neural network comprises the following steps:

[0010] Step 1: Obtain data: Calculate the power flow by simulating different operating modes and state fluctuations to obtain the original data as the input feature matrix;

[0011] Step 2: Data normalization: Perform Z-score normalization on the feature matrix to convert the dimensional original data into a dimensionless expression form X;

[0012] Step 3: Dataset division: Divide the data after the standardization process in step 2 into a training set for learning model parameters and a test set for the final evaluation of the model generalization ability;

[0013] Step 4: Construct a topology map and a feature map: Construct a topology map based on the topological structure of the power grid, and construct a feature map based on the feature matrix;

[0014] Step 5: Build a power grid operation model based on an adaptive multi-view graph convolutional neural network, including defining the graph convolution layer, acquiring multi-view embeddings, implementing the attention mechanism, and acquiring the target loss value. The topology graph and feature graph from step 4 are combined to build a specific graph convolution module and a common graph convolution module.

[0015] Step 6: Model training: Determine the optimizer parameter sink and model parameters, and use the training set to train the power grid operation model;

[0016] Step 7. Result analysis: After completing the model training, the grid operation model is run on the test set to predict the transformer short-circuit current, and the evaluation index is calculated to measure the model performance to evaluate the accuracy of the model in predicting the short-circuit current.

[0017] Furthermore, the characteristic matrix records the characteristics of the transformer bus before the short circuit fault, including bus voltage, phase angle, active power and reactive power injection.

[0018] Furthermore, the topology diagram in step 4 is used to describe the degree of connectivity between nodes and line impedance.

[0019] Furthermore, when constructing the feature graph based on the feature matrix in step 4, the similarity between each pair of nodes in the graph is first calculated. The similarity is achieved through cosine similarity or Gaussian kernel function. Then, for each node, multiple nodes that are most similar to it are selected as neighbors based on the level of similarity.

[0020] Furthermore, step 5 constructs a power grid operation model based on an adaptive multi-view graph convolutional neural network, which specifically includes:

[0021] (1) First, define the graph convolution layer: specify the input feature dimension and output feature dimension; initialize a weight matrix to map from the input feature space to the output feature space; in the forward propagation, multiply the adjacency matrix and the feature matrix to collect the local area information of each node; use the weight matrix to perform a linear transformation on the aggregated features, and then use a nonlinear activation function to enhance the nonlinear expression ability of the model;

[0022] (2) Obtaining embeddings from three different perspectives: First, based on the topology map and feature map, three embeddings are defined as input: topology-based embedding, feature-space-based embedding, and common embedding. Graph convolution operations are performed on the topology map and feature map, respectively, to obtain topology-based embedding and feature-space-based embedding. A common convolution module is introduced to process the topology-based and feature-map-based embedding information through a parameter sharing strategy, and the common information shared by the two spaces is extracted as a common embedding to enhance the power grid operation model's ability to capture common features.

[0023] (3) Attention mechanism implementation: First, a shared attention vector q is initialized to calculate the raw attention score of each node on different embeddings, which is done by the dot product of the nonlinear transformation of the node embedding and q; then, the raw attention score is normalized using the softmax function to ensure that the attention weight of each node is equal to 1 when added to the three embeddings, thus obtaining the final weight of each embedding; the weighted embeddings are then merged to form the final embedding representation of each node;

[0024] (4) Obtaining the target loss value: The power grid operation model introduces consistency constraints and difference constraints to improve the expressiveness and generalization capabilities of the embedding. The consistency constraint ensures the consistency of the embedding vectors obtained from different perspectives through L2 normalization and the Frobenius norm difference of the similarity matrix; the difference constraint encourages the model to capture the diversity in the feature space and topological space, and improves the model's ability to distinguish nodes of different categories; the consistency constraint and difference constraint are integrated in the loss function to obtain the final loss value.

[0025] A transformer short-circuit current evaluation device based on an adaptive multi-view graph neural network comprises:

[0026] The data acquisition module is used to obtain raw data as input feature matrix by simulating different operation modes and state fluctuations for power flow calculation;

[0027] A data standardization processing module is used to perform Z-score standardization on the feature matrix and convert the dimensional original data into a dimensionless expression form X;

[0028] The data set partitioning module is used to divide the data after the data standardization processing module is standardized into a training set for learning model parameters and a test set for ultimately evaluating the generalization ability of the model;

[0029] A topology map and feature map construction module is used to construct a topology map according to the topological structure of the power grid and to construct a feature map based on the feature matrix;

[0030] The power grid operation model construction module is used to build a power grid operation model based on an adaptive multi-view graph convolutional neural network. This includes defining graph convolution layers, obtaining multi-view embeddings, implementing the attention mechanism, and obtaining target loss values. It combines topology maps and feature maps to construct specific graph convolution modules and common graph convolution modules.

[0031] The model training module is used to determine the optimizer parameter sink and the model parameters, and train the power grid operation model using the training set;

[0032] The result analysis module is used to run the power grid operation model on the test set to predict the transformer short-circuit current after completing the model training, and calculate the evaluation index to measure the model performance to evaluate the accuracy of the model in predicting the short-circuit current.

[0033] Furthermore, the characteristic matrix records the characteristics of the transformer bus before the short circuit fault, including bus voltage, phase angle, active power and reactive power injection.

[0034] Furthermore, the power grid operation model construction module constructs a power grid operation model based on an adaptive multi-view graph convolutional neural network, specifically including:

[0035] (1) First, define the graph convolution layer: specify the input feature dimension and output feature dimension; initialize a weight matrix to map from the input feature space to the output feature space; in the forward propagation, multiply the adjacency matrix and the feature matrix to collect the local area information of each node; use the weight matrix to perform a linear transformation on the aggregated features, and then use a nonlinear activation function to enhance the nonlinear expression ability of the model;

[0036] (2) Obtaining embeddings from three different perspectives: First, based on the topology map and feature map, three embeddings are defined as input: topology-based embedding, feature-space-based embedding, and common embedding. Graph convolution operations are performed on the topology map and feature map, respectively, to obtain topology-based embedding and feature-space-based embedding. A common convolution module is introduced to process the topology-based and feature-map-based embedding information through a parameter sharing strategy, and the common information shared by the two spaces is extracted as a common embedding to enhance the power grid operation model's ability to capture common features.

[0037] (3) Attention mechanism implementation: First, a shared attention vector q is initialized to calculate the raw attention score of each node on different embeddings, which is done by the dot product of the nonlinear transformation of the node embedding and q; then, the raw attention score is normalized using the softmax function to ensure that the attention weight of each node is equal to 1 when added to the three embeddings, thus obtaining the final weight of each embedding; the weighted embeddings are then merged to form the final embedding representation of each node;

[0038] (4) Obtaining the target loss value: The power grid operation model introduces consistency constraints and difference constraints to improve the expressiveness and generalization capabilities of the embedding. The consistency constraint ensures the consistency of the embedding vectors obtained from different perspectives through L2 normalization and the Frobenius norm difference of the similarity matrix; the difference constraint encourages the model to capture the diversity in the feature space and topological space, and improves the model's ability to distinguish nodes of different categories; the consistency constraint and difference constraint are integrated in the loss function to obtain the final loss value.

[0039] A transformer short-circuit current assessment system based on an adaptive multi-view graph neural network comprises: a computer-readable storage medium and a processor;

[0040] The computer-readable storage medium is used to store executable instructions;

[0041] The processor is used to read the executable instructions stored in the computer-readable storage medium and execute the transformer short-circuit current evaluation method based on the adaptive multi-view graph neural network.

[0042] A non-transitory computer-readable storage medium stores a computer program, which, when executed by a processor, implements the transformer short-circuit current evaluation method based on an adaptive multi-view graph neural network.

[0043] The present invention has the following beneficial effects:

[0044] 1. Effective integration of node features and topological structure information: This invention significantly improves the accuracy of short-circuit current assessment by introducing specific graph convolution modules and public graph convolution modules. The specific graph convolution modules operate independently on the topological graph and feature graph respectively, extracting their respective unique information, thereby ensuring that the model can fully understand the complex structure of the power grid. The public graph convolution module adopts a parameter sharing strategy, allowing the extraction of shared features on the two graphs, enhancing the grasp of the commonality of the information space. This dual processing mechanism ensures the efficient integration of node features and topological structure information, thereby improving the predictive ability of the model.

[0045] 2. Design an adaptive mechanism that enables the model to dynamically adjust weights between different feature representations to optimize short-circuit current prediction. Specifically, the model utilizes an attention mechanism to dynamically fuse node topological features and feature information. By generating specific and shared embeddings, the model calculates attention weights and applies them to embedding fusion, achieving adaptive integration of information. This innovation enables the model to automatically adjust feature importance based on different input conditions, thereby improving prediction flexibility and accuracy.

[0046] 3. Constructing a multi-perspective graph neural network: This paper constructs a multi-perspective graph neural network that aims to capture the complex relationships between nodes in the power grid and learn richer feature representations. The model not only considers traditional topological graphs, but also introduces feature graphs based on feature matrices. By performing graph convolution operations in both topological space and feature space, feature learning is performed from multiple perspectives. This multi-perspective design enables the model to fully understand the dynamic behavior of the power grid and enhances the ability to distinguish different types of nodes, thereby improving the accuracy and robustness of short-circuit current assessment. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 This is a flow chart of a transformer short-circuit current evaluation method based on an adaptive multi-view graph neural network according to an embodiment of the present invention;

[0048] Figure 2 is a schematic diagram of a power grid operation model according to an embodiment of the present invention;

[0049] Figure 3 is a trend diagram showing the change of the loss value as the number of iterations increases according to an embodiment of the present invention;

[0050] Figure 4 is the error between the predicted short-circuit current and the actual short-circuit current in the fault node according to the embodiment of the present invention;

[0051] Figure 5 This is a clustering trend diagram of an embodiment of the present invention. DETAILED DESCRIPTION

[0052] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0053] The inventors of this application discovered through research during the process of realizing the present invention that: when evaluating transformer short-circuit current, traditional GCN mainly focuses on the topology of the system itself, that is, the adjacency relationship between nodes. However, relying solely on the topology for graph convolution operations will result in a large amount of valuable information being lost. Because this single perspective ignores the feature graph of the node, the characteristics of the node not only include the intrinsic properties of the node, but also contain the implicit relationship between nodes, that is, the feature graph. These feature graphs are of great significance for understanding the dynamic behavior of the power grid and predicting short-circuit current. Based on multiple perspectives, the model can better understand the geometric relationship and directional similarity between nodes. This makes the model more flexible and comprehensive in capturing complex relationships and feature patterns, and the multi-perspective structure can enhance the robustness of the model.

[0054] With a single adjacency matrix, the performance of graph network models is easily affected by the graph structure, especially in the presence of noisy or incomplete graph data. By combining different adjacency matrices, the model can perform a weighted average of multiple relationships, thereby reducing the overall performance degradation caused by inaccuracies in a particular channel. In addition, multiple perspectives can also improve the model's expressiveness. Different adjacency matrices represent different graph structures and relationship types. By performing convolution operations on these structures in parallel, the model can learn richer feature representations.

[0055] Furthermore, traditional methods for merging multi-view features fail to fully exploit the differences in information captured by each channel. Fixed-weight averaging or simple combination fails to dynamically adjust the importance of different feature representations, potentially leading to information loss or confusion. For example, in certain tasks, node feature topology may be more important than structural topology, and vice versa. However, fixed weighting methods fail to recognize and exploit this difference, potentially diluting the contribution of a particular important feature, thus impacting the model's overall performance. However, methods based on adaptive multi-view feature fusion allow the model to dynamically adjust the weights of each feature representation during learning, enabling the model to adaptively focus on the most relevant feature representations. This capability enhances the model's flexibility and generalization capabilities when processing diverse graph data. Furthermore, adaptive learning can capture the interrelationships and relative importance of different feature representations, thereby improving the model's ability to recognize complex patterns. For example, under certain conditions, an adjacency matrix based on node features may better represent target information than other adjacency matrices. Therefore, the attention mechanism can assign higher weight to this matrix, thereby optimizing the final prediction results.

[0056] See also Figure 1 The embodiment of the present invention provides a transformer short-circuit current evaluation method based on an adaptive multi-view graph neural network, comprising the following steps:

[0057] Step 1. Data acquisition: The raw data is obtained by simulating different operating modes and state fluctuations to perform power flow calculations. The raw data is used as the input feature matrix. The feature matrix records the characteristics of the transformer bus before the short-circuit fault, including bus voltage, phase angle, active power, and reactive power injection.

[0058] Step 2. Data Normalization: Perform Z-score normalization on the feature matrix to convert the dimensional raw data into a dimensionless representation, X. Data normalization adjusts features to the same scale, eliminating dimensionality effects, thereby accelerating model convergence, improving algorithm stability and performance, and ensuring fair performance across different features.

[0059] Step 3. Dataset division: After completing data standardization, divide the dataset into two parts according to a certain ratio: a training set, which is used to learn model parameters; and a test set, which is used to finally evaluate the generalization ability of the model.

[0060] Step 4. Constructing a topology map and feature map: A topology map is constructed based on the grid's topology. The topology map describes the connectivity between nodes and the line impedance. A feature map is constructed based on the feature matrix. To construct the feature map, the similarity between each pair of nodes in the graph is first calculated, typically using cosine similarity or a Gaussian kernel function. Next, for each node, the k most similar nodes are selected as neighbors based on the similarity.

[0061] Step 5. Construct a power grid operation model based on an adaptive multi-view graph convolutional neural network: including defining the graph convolution layer, obtaining multi-view embedding values, implementing the attention mechanism, and obtaining the target loss. In particular, the topology map and feature map of step 4 are combined to construct a specific graph convolution module and a common graph convolution module. Figure 2 As shown, the specific implementation steps of step 5 are as follows:

[0062] (1) First, define the graph convolutional layer: specify the input feature dimension and the output feature dimension; initialize a weight matrix to map from the input feature space to the output feature space. In the forward propagation, the adjacency matrix and the feature matrix are multiplied to collect the local area information of each node. The weight matrix is used to linearly transform the aggregated features, and then a nonlinear activation function is used to enhance the nonlinear expression ability of the model.

[0063] (2) Obtaining embeddings from three different perspectives: First, based on the topology map and feature map, three embeddings are defined as input: topology-based embedding, feature-space-based embedding, and common embedding. Graph convolution operations are performed on the topology map and feature map, respectively, to obtain topology-based embedding and feature-space-based embedding. A common convolution module is introduced to process the topology-based and feature-map-based embedding information through a parameter sharing strategy, extracting the common information shared by the two spaces as a common embedding to enhance the power grid operation model's ability to capture common features.

[0064] (3) Attention mechanism implementation: First, a shared attention vector q is initialized to calculate the raw attention scores of each node on different embeddings, which is usually done by a nonlinear transformation of the node embedding and the dot product of q. Then, these raw scores are normalized using the softmax function to ensure that the attention weight of each node is equal to 1 when added to the three embeddings, thereby obtaining the final weight of each embedding. These weighted embeddings are then merged to form the final embedding representation of each node.

[0065] (4) Finally, the target loss value is obtained: In order to improve the expressiveness and generalization ability of the embedding, the power grid operation model introduces consistency constraints and difference constraints. The consistency constraint ensures the consistency of the embedding vectors obtained from different perspectives through L2 normalization and the Frobenius norm difference of the similarity matrix. The difference constraint encourages the model to capture the diversity in the feature space and topological space, and improves the model's ability to distinguish nodes of different categories; since the evaluation of short-circuit current is a regression task, this embodiment selects MSE as the measure of the loss between the predicted value and the actual value. The consistency constraint and difference constraint are then integrated into the loss function to obtain the final loss value.

[0066] Step 6. Model Training: After building a power grid operation model based on an adaptive multi-view graph convolutional neural network, an optimization algorithm is used to adjust model parameters based on the calculated gradients using the training set to minimize the loss function. In the optimizer, important hyperparameters are designed, such as the learning rate, momentum, and weight decay. The learning rate controls the magnitude of model parameter updates and determines the step size of the parameter update in each iteration. A higher learning rate means a larger parameter update step, which may lead to faster training convergence, but may also skip the optimal solution or cause training instability. A lower learning rate means a smaller parameter update step. While this allows the model to more stably approach the optimal solution, it may result in very slow training. Weight decay is a regularization technique that suppresses model complexity and prevents overfitting by adding a penalty term based on the sum of squared weights to the loss function. Weight decay is essentially equivalent to L2 regularization.

[0067] Step 7. Result Analysis: After completing model training, make predictions on the test set and calculate evaluation metrics such as mean squared error to measure model performance. Figure 3 The figure depicts the trend of loss value as the number of iterations increases. It can be found that as the number of iterations increases, the loss value continues to decrease until it reaches a convergence state. Figure 4 is the error between the predicted short-circuit current and the actual short-circuit current in the fault node. As can be seen from the figure, all errors are less than 0.06, which is within the acceptable range. Figure 5 In the CNN dataset, the features learned by the model were reduced in dimension and visualized, showing a clear clustering trend. This demonstrates that the CNN model can effectively capture the inherent structure in the data and cluster similar instances together, thereby improving the performance of classification or other downstream tasks.

[0068] The present invention has the following characteristics:

[0069] 1. Adaptive Multi-View Graph Neural Network Architecture: This paper proposes a transformer short-circuit current estimation method based on an adaptive multi-view graph neural network. The core of this method lies in its ability to simultaneously consider node characteristics and grid topology. By constructing different adjacency matrices, it captures the complex relationships between nodes in the grid, thereby improving the accuracy of short-circuit current prediction.

[0070] 2. Feature Fusion and Adaptive Weight Adjustment: This paper employs an adaptive feature fusion approach, allowing the model to dynamically adjust the weights of different feature representations during the learning process to optimize short-circuit current prediction. This adaptive mechanism captures the interrelationships and relative importance of different feature representations, improving the model's ability to recognize complex patterns.

[0071] 3. Optimization Objectives with Consistency and Difference Constraints: During the model optimization process, this paper introduces consistency and difference constraints to ensure that the embedding vectors obtained from different perspectives are consistent and capture the different characteristics of nodes in feature space and topological space. This design helps improve the model's generalization ability and the ability to distinguish different types of nodes.

[0072] An embodiment of the present invention further provides a transformer short-circuit current evaluation device based on an adaptive multi-view graph neural network, comprising:

[0073] The data acquisition module is used to obtain raw data as input feature matrix by simulating different operation modes and state fluctuations for power flow calculation;

[0074] A data standardization processing module is used to perform Z-score standardization on the feature matrix and convert the dimensional original data into a dimensionless expression form X;

[0075] The data set partitioning module is used to divide the data after the data standardization processing module is standardized into a training set for learning model parameters and a test set for ultimately evaluating the generalization ability of the model;

[0076] A topology map and feature map construction module is used to construct a topology map according to the topological structure of the power grid and to construct a feature map based on the feature matrix;

[0077] The power grid operation model construction module is used to build a power grid operation model based on an adaptive multi-view graph convolutional neural network. This includes defining graph convolution layers, obtaining multi-view embeddings, implementing the attention mechanism, and obtaining target loss values. It combines topology maps and feature maps to construct specific graph convolution modules and common graph convolution modules.

[0078] The model training module is used to determine the optimizer parameter sink and the model parameters, and train the power grid operation model using the training set;

[0079] The result analysis module is used to run the power grid operation model on the test set to predict the transformer short-circuit current after completing the model training, and calculate the evaluation index to measure the model performance to evaluate the accuracy of the model in predicting the short-circuit current.

[0080] Another embodiment of the present invention provides a transformer short-circuit current assessment system based on an adaptive multi-view graph neural network, comprising: a computer-readable storage medium and a processor;

[0081] The computer-readable storage medium is used to store executable instructions;

[0082] The processor is used to read the executable instructions stored in the computer-readable storage medium and execute the transformer short-circuit current evaluation method based on the adaptive multi-view graph neural network.

[0083] Another embodiment of the present invention provides a non-transitory computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the method for evaluating transformer short-circuit current based on an adaptive multi-view graph neural network is implemented.

[0084] Those skilled in the art will appreciate that the embodiments of the present application can be provided as methods, systems, or computer program products. Therefore, the present application can adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application can adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) that contain computer-usable program code.

[0085] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the steps in the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0086] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0087] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 The steps for the function specified in one or more boxes.

[0088] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents. Any modification or equivalent replacement that does not depart from the spirit and scope of the present invention should be covered by the scope of protection of the claims of the present invention.

Claims

1. A transformer short-circuit current assessment method based on an adaptive multi-view graph neural network, characterized in that: The steps include: Step 1: Obtain data: Calculate the power flow by simulating different operating modes and state fluctuations to obtain the original data as the input feature matrix; Step 2: Data normalization: Perform Z-score normalization on the feature matrix to convert the dimensional original data into a dimensionless expression form X; Step 3: Dataset division: Divide the data after the standardization process in step 2 into a training set for learning model parameters and a test set for the final evaluation of the model generalization ability; Step 4: Construct a topology map and a feature map: Construct a topology map based on the topological structure of the power grid, and construct a feature map based on the feature matrix; Step 5: Build a power grid operation model based on an adaptive multi-view graph convolutional neural network, including defining the graph convolution layer, acquiring multi-view embeddings, implementing the attention mechanism, and acquiring the target loss value. The topology graph and feature graph from step 4 are combined to build a specific graph convolution module and a common graph convolution module. Step 6: Model training: Determine the optimizer parameter sink and model parameters, and use the training set to train the power grid operation model; Step 7. Result analysis: After completing the model training, the grid operation model is run on the test set to predict the transformer short-circuit current, and the evaluation index is calculated to measure the model performance to evaluate the accuracy of the model in predicting the short-circuit current.

2. The transformer short-circuit current evaluation method based on adaptive multi-view graph neural network according to claim 1, characterized in that: The characteristic matrix records the characteristics of the transformer bus before a short circuit fault, including bus voltage, phase angle, active power and reactive power injection.

3. The transformer short-circuit current evaluation method based on adaptive multi-view graph neural network according to claim 1, characterized in that: The topology diagram in step 4 is used to describe the degree of connectivity between nodes and the line impedance.

4. The transformer short-circuit current evaluation method based on adaptive multi-view graph neural network according to claim 1, characterized in that: When constructing the feature graph based on the feature matrix in step 4, the similarity between each pair of nodes in the graph is first calculated. The similarity is achieved through cosine similarity or Gaussian kernel function. Then, for each node, the most similar nodes are selected as neighbors based on the similarity.

5. The transformer short-circuit current evaluation method based on adaptive multi-view graph neural network according to claim 1, characterized in that: Step 5 builds a power grid operation model based on an adaptive multi-view graph convolutional neural network, which specifically includes: (1) First, define the graph convolution layer: specify the input feature dimension and output feature dimension; initialize a weight matrix to map from the input feature space to the output feature space; in the forward propagation, multiply the adjacency matrix and the feature matrix to collect the local area information of each node; use the weight matrix to perform a linear transformation on the aggregated features, and then use a nonlinear activation function to enhance the nonlinear expression ability of the model; (2) Obtaining embeddings from three different perspectives: First, based on the topology map and feature map, three embeddings are defined as input: topology-based embedding, feature-space-based embedding, and common embedding. Graph convolution operations are performed on the topology map and feature map, respectively, to obtain topology-based embedding and feature-space-based embedding. A common convolution module is introduced to process the topology-based and feature-map-based embedding information through a parameter sharing strategy, and the common information shared by the two spaces is extracted as a common embedding to enhance the power grid operation model's ability to capture common features. (3) Attention mechanism implementation: First, a shared attention vector q is initialized to calculate the raw attention score of each node on different embeddings, which is done by the dot product of the nonlinear transformation of the node embedding and q; then, the raw attention score is normalized using the softmax function to ensure that the attention weight of each node is equal to 1 when added to the three embeddings, thus obtaining the final weight of each embedding; the weighted embeddings are then merged to form the final embedding representation of each node; (4) Obtaining the target loss value: The power grid operation model introduces consistency constraints and difference constraints to improve the expressiveness and generalization capabilities of the embedding. The consistency constraint ensures the consistency of the embedding vectors obtained from different perspectives through L2 normalization and the Frobenius norm difference of the similarity matrix; the difference constraint encourages the model to capture the diversity in the feature space and topological space, and improves the model's ability to distinguish nodes of different categories; the consistency constraint and difference constraint are integrated in the loss function to obtain the final loss value.

6. A transformer short-circuit current evaluation device based on an adaptive multi-view graph neural network, characterized in that: include: The data acquisition module is used to obtain raw data as input feature matrix by simulating different operation modes and state fluctuations for power flow calculation; A data standardization processing module is used to perform Z-score standardization on the feature matrix and convert the dimensional original data into a dimensionless expression form X; The data set partitioning module is used to divide the data after the data standardization processing module is standardized into a training set for learning model parameters and a test set for ultimately evaluating the generalization ability of the model; A topology map and feature map construction module is used to construct a topology map according to the topological structure of the power grid and to construct a feature map based on the feature matrix; The power grid operation model construction module is used to build a power grid operation model based on an adaptive multi-view graph convolutional neural network. This includes defining graph convolution layers, obtaining multi-view embeddings, implementing the attention mechanism, and obtaining target loss values. It combines topology maps and feature maps to construct specific graph convolution modules and common graph convolution modules. The model training module is used to determine the optimizer parameter sink and the model parameters, and train the power grid operation model using the training set; The result analysis module is used to run the power grid operation model on the test set to predict the transformer short-circuit current after completing the model training, and calculate the evaluation index to measure the model performance to evaluate the accuracy of the model in predicting the short-circuit current.

7. The transformer short-circuit current evaluation device based on adaptive multi-view graph neural network according to claim 6, characterized in that: The characteristic matrix records the characteristics of the transformer bus before a short circuit fault, including bus voltage, phase angle, active power and reactive power injection.

8. The transformer short-circuit current evaluation device based on adaptive multi-view graph neural network according to claim 6, characterized in that: The power grid operation model construction module constructs a power grid operation model based on an adaptive multi-view graph convolutional neural network, specifically including: (1) First, define the graph convolution layer: specify the input feature dimension and output feature dimension; initialize a weight matrix to map from the input feature space to the output feature space; in the forward propagation, multiply the adjacency matrix and the feature matrix to collect the local area information of each node; use the weight matrix to perform a linear transformation on the aggregated features, and then use a nonlinear activation function to enhance the nonlinear expression ability of the model; (2) Obtaining embeddings from three different perspectives: First, based on the topology map and feature map, three embeddings are defined as input: topology-based embedding, feature-space-based embedding, and common embedding. Graph convolution operations are performed on the topology map and feature map, respectively, to obtain topology-based embedding and feature-space-based embedding. A common convolution module is introduced to process the topology-based and feature-map-based embedding information through a parameter sharing strategy, and the common information shared by the two spaces is extracted as a common embedding to enhance the power grid operation model's ability to capture common features. (3) Attention mechanism implementation: First, a shared attention vector q is initialized to calculate the raw attention score of each node on different embeddings, which is done by the dot product of the nonlinear transformation of the node embedding and q; then, the raw attention score is normalized using the softmax function to ensure that the attention weight of each node is equal to 1 when added to the three embeddings, thus obtaining the final weight of each embedding; the weighted embeddings are then merged to form the final embedding representation of each node; (4) Obtaining the target loss value: The power grid operation model introduces consistency constraints and difference constraints to improve the expressiveness and generalization capabilities of the embedding. The consistency constraint ensures the consistency of the embedding vectors obtained from different perspectives through L2 normalization and the Frobenius norm difference of the similarity matrix; the difference constraint encourages the model to capture the diversity in the feature space and topological space, and improves the model's ability to distinguish nodes of different categories; the consistency constraint and difference constraint are integrated in the loss function to obtain the final loss value.

9. A transformer short-circuit current assessment system based on an adaptive multi-view graph neural network, comprising: Computer-readable storage media and processor; The computer-readable storage medium is used to store executable instructions; The processor is used to read the executable instructions stored in the computer-readable storage medium and execute the transformer short-circuit current evaluation method based on adaptive multi-view graph neural network according to any one of claims 1 to 5.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, the method for evaluating transformer short-circuit current based on an adaptive multi-view graph neural network according to any one of claims 1 to 5 is implemented.

Images