Heterogeneous graph attention network-based power distribution network short circuit fault diagnosis method and device, and medium

Abstract

The invention discloses a power distribution network short circuit fault diagnosis method and device based on a heterogeneous graph attention network and a medium, and belongs to the field of power system protection and control, and the method comprises the steps: building a heterogeneous power distribution network topological graph structure, and measuring the electric data of each node in the power distribution network topological graph; preprocessing the measured electric data and constructing a fault data set; constructing a power distribution network short-circuit fault diagnosis model based on the heterogeneous graph attention network; defining a training hyper-parameter and a loss function of the power distribution network short-circuit fault diagnosis model, and training the power distribution network short-circuit fault diagnosis model; and when a power distribution network fault occurs, preprocessing real-time measurement data, and inputting the data into the optimal power distribution network short-circuit fault diagnosis model to obtain prediction results of a node fault state, a fault category and a fault phase. According to the method, the fault of the power distribution network can be accurately and effectively diagnosed under the condition of distributed power supply deployment, and the method has relatively high diagnosis accuracy and robustness.

Description

Technical Field

[0001] The present invention belongs to the field of power system protection and control, and specifically relates to a distribution network short-circuit fault diagnosis method, equipment and medium based on a heterogeneous graph attention network. Background Art

[0002] Fault diagnosis is a critical task in the operation and maintenance of power systems, especially in distribution networks, which are particularly important due to their complex interconnectedness and large scale. Failure to take appropriate measures in a timely manner during a fault event can lead to cascading power outages in the distribution system, resulting in serious power outages. To ensure the continued stable operation of the power grid when a fault occurs, it is necessary not only to accurately locate the fault location but also to fully identify the fault mode, including key characteristics such as the short-circuit type and short-circuit phase. This information can help grid operators quickly locate the source of the fault and develop precise repair plans, thereby minimizing the impact of the fault.

[0003] Traditional fault diagnosis methods typically rely on statistical measurements, signal processing, and the physical characteristics of fault events. While effective to a certain extent, these approaches often rely heavily on expert experience and have high computational complexity, making them difficult to adapt to the dynamic and complex environment of modern power grids. In recent years, with the rapid development of deep learning technology, most fault diagnosis systems have begun to shift towards data-driven approaches. While these data-driven approaches vary widely in type and architecture, they can be broadly categorized into two types: those based on graph neural networks (GNNs), and those that integrate multi-layer perceptrons (MLPs) with convolutional neural networks. While these methods have achieved promising results in specific scenarios, there remains a lack of solutions that balance model generalization and the integration of multi-source heterogeneous data to address common engineering challenges in real-world deployments, such as measurement noise interference and protection device timing deviations. Furthermore, many studies rely solely on theoretical assumptions and fail to fully consider the complexities of real-world applications.

[0004] Therefore, conducting research on distribution network fault diagnosis is of great significance to improving the operational reliability and safety of new power systems. It is urgent to develop a more accurate and adaptable fault diagnosis technology to reduce power outage losses caused by faults, achieve a safer, more reliable and efficient power supply, assist in the transformation of energy structure, and support the sustainable development goals of society. Summary of the Invention

[0005] In response to the deficiencies in the prior art, the present invention utilizes multi-source data and a heterogeneous graph attention network (HGAT) to provide a distribution network short-circuit fault diagnosis method, device, and medium based on a heterogeneous graph attention network, aiming to address the limitations of existing fault location technologies and meet the complex needs in practical applications. After a fault occurs, the designed fault diagnosis model can not only quickly determine the fault section, but also accurately identify the fault category and the affected phase. In particular, for faults in complex situations such as distribution network topology reconstruction and active distribution networks, the present invention can achieve more accurate fault diagnosis and significantly improve the efficiency and accuracy of distribution network fault handling. This method not only makes up for the shortcomings of traditional technologies, but also meets the needs of modern distribution networks for efficient and accurate fault diagnosis.

[0006] To achieve the above object, the present invention adopts the following technical solutions:

[0007] A distribution network short circuit fault diagnosis method based on heterogeneous graph attention network includes the following steps:

[0008] Establish a heterogeneous distribution network topology structure and measure the electrical data of each node in the distribution network topology;

[0009] Preprocess the measured electrical data and construct a fault dataset;

[0010] Construct a distribution network short circuit fault diagnosis model based on heterogeneous graph attention network;

[0011] Define the training hyperparameters and loss function of the distribution network short-circuit fault diagnosis model. Divide the fault dataset into training, validation, and test sets. Use the training and test sets to train the distribution network short-circuit fault diagnosis model. Fine-tune the parameters during training. Use the validation set to verify the performance of the trained distribution network short-circuit fault diagnosis model and obtain the optimal distribution network short-circuit fault diagnosis model.

[0012] When a distribution network fault occurs, the real-time measurement data is preprocessed and input into the optimal distribution network short-circuit fault diagnosis model to obtain the prediction results of the node fault state, fault type and fault phase.

[0013] To optimize the above technical solutions, specific measures taken also include:

[0014] Furthermore, the establishment of a heterogeneous distribution network topology structure is specifically as follows:

[0015] The two types of measuring devices in the distribution network, intelligent distribution transformer terminals (TTUs) and feeder terminal units (FTUs), are used as nodes in the distribution network topology. The power transmission relationships between the measuring devices are used as edges in the distribution network topology. Based on the distance from the substation, the nodes are numbered from far to near, from the main line to the branch, and first the TTU nodes and then the FTU nodes to construct a distribution network topology diagram. The distribution network topology diagram satisfies the requirement that the sum of the node and edge types is greater than 2.

[0016] Furthermore, the electrical data of each node in the distribution network topology diagram is specifically measured as follows:

[0017] For the intelligent distribution transformer terminal TTU node, the measurement data obtained include three-phase voltage, three-phase current, zero-sequence voltage and zero-sequence voltage; for the feeder terminal unit FTU node, the measurement data obtained include three-phase voltage, three-phase current, zero-sequence voltage, zero-sequence voltage and fault current direction; among them, the three-phase voltage, three-phase current, zero-sequence voltage and zero-sequence voltage are the amplitude data within the measurement time period, and the value of the fault current direction data is "1", "0" or "-1", which respectively indicates that the fault current direction is from the bus to the line, no fault current is detected and the fault current direction is from the line to the bus.

[0018] Furthermore, the preprocessing of the measured electrical data is specifically as follows:

[0019] The missing parts of the measured electrical data are filled with zero values ​​to fill the missing parts, and Z-score or Min-Max normalization is performed on them.

[0020] Furthermore, the construction of the fault data set is specifically as follows:

[0021] Get the node information set X; for a graph structure with n nodes, the node set V in the graph is V = {v1, v2, ..., v n}, the set of node information X is X={x1,x2,…,x n}, for each node's node information x i ,i∈[1,n], which is composed of node type x node , node number x ID , node feature x feat with node label x lable Together, they constitute x i =(x node ,x ID ,x feat ,x lable );

[0022] Get the edge information set E, for the edge information set E, E=(e1,e2,…,e m ), where m represents the number of edges, and the edge information e for each edgej ,j∈[1,m], which is composed of the starting node number, the target node number and the edge type, that is, e j =(v ID_start ,v ID_to ,e type ), where edge type e type There are three categories: represents the edge from FTU node to TTU node; Represents the edge from TTU node to FTU node; represents the edge from TTU node to TTU node;

[0023] Define a heterogeneous graph element path Φ and construct a fault information label L. The heterogeneous graph element path Φ is composed of multiple groups of edge types. According to the actual situation of the distribution network, a possible element path is constructed by splicing two groups of edge types, as follows:

[0024] The specific path of the first meta-path Φ1 is: from FTU node to TTU node and then to TTU node:

[0025] The specific path of the second meta-path Φ2 is: from TTU node to FTU node and then to TTU node:

[0026] The specific path of the third element path Φ3 is: from TTU node to TTU node and then to FTU node:

[0027] The specific path of the fourth element path Φ4 is: from TTU node to TTU node and then to TTU node:

[0028] The fault information tag is composed of L=(L location ,L type ,L phase ), where the fault status label L location For fault location tasks, the fault status label is defined as follows: when a short circuit fault occurs on a line, the fault status label of the node not affected by the short circuit fault is "1", and the fault status label of the node affected by the short circuit fault is "0";

[0029] Fault category label L type It is used for fault classification tasks. The fault categories include no fault, single-phase grounding fault, two-phase interphase short circuit fault, two-phase short circuit grounding fault, three-phase short circuit fault, and three-phase short circuit grounding fault. The fault category labels are mapped into 6-dimensional binary vectors using one-hot encoding.

[0030] Fault phase label L phaseIt is used for fault phase classification tasks. Fault phases include no fault phase, A phase fault, B phase fault, C phase fault, AB two-phase fault, BC two-phase fault, CA two-phase fault, and ABC three-phase fault. The fault phase labels are encoded using a one-hot encoding method, mapping the eight fault phases into an 8-dimensional binary vector.

[0031] Furthermore, the construction of a distribution network short circuit fault diagnosis model based on heterogeneous graph attention network is specifically as follows:

[0032] The distribution network short-circuit fault diagnosis model includes a fault location model and a fault type and fault phase classification model;

[0033] The fault localization model uses two layers of heterogeneous graph attention networks (HGALs) as feature encoders to capture features of different fault locations. A multi-layer perceptron (MLP) is used as the output layer, and Sigmoid is used as the activation function of the final output layer. The output layer outputs the probability that each node is in a fault state. A skip connection layer is introduced between the input of the fault localization model and the second HGAL layer to directly transfer low-level features to high-level features through dimensional mapping.

[0034] The fault category and fault phase classification model uses a two-layer heterogeneous graph attention network (HGAL) as a feature encoder to capture the features of different fault categories and fault phases. A global average pooling layer is introduced before the output layer to average all node features and aggregate them into features for the entire graph. A fully connected layer (FCL) is used as the output layer, and Softmax is used as the activation function of the output layer. The output layer outputs the probability distribution of the fault category and fault phase for different categories.

[0035] A skip connection layer is introduced between the input of the fault category and fault phase classification model and the second HGAL layer to directly transfer low-level features to high-level features through dimensional mapping.

[0036] Furthermore, the loss function of the distribution network short-circuit fault diagnosis model adopts a focal loss function with an adaptive modulation factor, and the specific expression is as follows:

[0037]

[0038] Where, Loss FL is the focal loss, is the adaptive modulation factor, p i,k is the probability that the i-th sample belongs to category k; N is the total number of samples, K is the total number of categories, y i,k is a label used to mark whether the i-th sample belongs to category k. If it does, y i,k =1, otherwise y i,k =0;γ init is the initial modulation factor set; Nk_samples is the total number of samples in the kth category; i iteration The number of iterations during model training;

[0039] During the parameter updating process, the Adam optimizer is used to perform adaptive adjustment of the parameters. The training ends when the loss function value of the distribution network short-circuit fault diagnosis model converges.

[0040] Furthermore, each heterogeneous graph attention network layer HGAL in the fault localization model includes two levels of attention, namely node-level attention and semantic-level attention; the node-level attention dynamically assigns different attention weights according to the importance of the node's neighbors under different meta-paths; after obtaining the feature representation of the node under different meta-paths through node-level attention, the semantic-level attention takes it as input and calculates a semantic-level attention weight for each meta-path to represent the importance of the meta-path to the target node, and obtains the final node feature by weighted aggregation of features of different meta-paths.

[0041] The present invention also proposes an electronic device, comprising: a memory, a processor, and a computer program stored in the memory and runnable on the processor. When the processor executes the computer program, it implements the distribution network short-circuit fault diagnosis method based on the heterogeneous graph attention network as described above.

[0042] The present invention also proposes a computer-readable storage medium storing a computer program, wherein the computer program enables a computer to execute the distribution network short circuit fault diagnosis method based on the heterogeneous graph attention network as described above.

[0043] The beneficial effects of the present invention are:

[0044] The present invention utilizes FTU-TTU multi-source data fusion under spatiotemporal synchronization constraints to realize distribution network fault diagnosis, and improves positioning accuracy and fault category identification reliability through multi-dimensional feature extraction and association analysis.

[0045] Fault diagnosis based on heterogeneous graph attention networks not only has more advantages in capturing the fault characteristics of complex systems, but also constructs heterogeneous topologies of distribution networks with different nodes. It not only further refines the topological structure of the distribution network and provides sufficient data support for fault diagnosis, but also promotes the full integration of multi-source data, providing richer data support for fault diagnosis.

[0046] The fault localization model uses a two-layer heterogeneous graph attention network (HGAL) as a feature encoder to capture features of different fault locations. A multi-layer perceptron (MLP) is used as the output layer, and a sigmoid activation function is used as the final output layer to output the probability of each node being in a faulty state. Furthermore, a skip connection layer is introduced between the model input and the second HGAL layer to improve the model's gradient propagation efficiency, mitigate degradation issues in deep networks, and enhance feature reuse, thereby improving the model's representation accuracy of fault location features and training stability.

[0047] The fault category and fault phase classification model uses a two-layer heterogeneous graph attention network (HGAL) as a feature encoder to capture the characteristics of different fault categories and fault distinctions. A global average pooling layer is introduced before the output layer to average all node features and aggregate them into features for the entire graph. A fully connected layer (FCL) serves as the output layer, using Softmax as the activation function to predict the output as the probabilities of different categories. Similarly, a skip connection layer is introduced between the model input and the second HGAL layer. Through dimensional mapping, low-level features are directly transferred to higher levels to avoid vanishing or exploding gradients during training.

[0048] The focus loss function with adaptive modulation factor designed by the present invention is adaptively adjusted by setting the introduction γ k * , it can be achieved that in the early stage of training when the number of iterations is small, the modulation factors of samples of different numbers are large, thereby accelerating the model's learning of complex features; as the number of iterations increases, in the later stage of training, the modulation factors of samples of different categories gradually decrease, so that the model can balance the contribution of all samples to model learning, avoid over-focusing on difficult samples and neglecting easy samples, and ensure that the overall performance of the model is improved. At the same time, to address the problem of class imbalance, the adaptive modulation factor can make the modulation factor of categories with fewer samples higher, so that the model can better learn the features in these categories; for categories with more samples, the value of the modulation factor can be appropriately reduced to reduce its impact on model training. In this way, the focal loss function with an adaptive modulation factor can effectively solve the problem of imbalance in the number of samples between categories while maintaining the focal loss's focus on difficult-to-classify samples, thereby improving the overall performance of the model in imbalanced data scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 This is a flow chart of the distribution network short-circuit fault diagnosis method based on heterogeneous graph attention network proposed in this invention.

[0050] Figure 2 This is an example diagram of distribution network topology.

[0051] Figure 3This is the structural diagram of the fault localization model based on heterogeneous graph attention network.

[0052] Figure 4 Schematic diagram of the process of aggregating features for node attention mechanism and semantic attention mechanism.

[0053] Figure 5 This is the structural diagram of the fault category and fault phase classification model based on heterogeneous graph attention network.

[0054] Figure 6 This is a diagram showing the changes in the adaptive modulation factor under different sample numbers.

[0055] Figure 7 is a visualization diagram of node features. Figure 7a For fault type identification, Figure 7b Identify the fault phase.

[0056] Figure 8 The following is a line chart comparing performance under different topology changes. DETAILED DESCRIPTION

[0057] The following will be combined with the drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0058] Example 1

[0059] This paper proposes a distribution network short circuit fault diagnosis method based on heterogeneous graph attention network. The overall process of the method is as follows: Figure 1 As shown, the following steps are included:

[0060] S1. Establish a heterogeneous distribution network topology structure and measure the electrical data of each node in the distribution network topology.

[0061] S1.1. Take the two types of measuring devices in the distribution network, intelligent distribution transformer terminal TTU and feeder terminal unit FTU, as nodes in the distribution network topology, and the power transmission relationship between each measuring device as the edge in the distribution network topology. Based on the distance from the substation, the nodes are numbered from far to near, from the main line to the branch, first the TTU node and then the FTU node to construct the distribution network topology. The distribution network topology satisfies the requirement that the sum of the node and edge types is greater than 2. Figure 2Taking the IEEE 33-node standard distribution network topology as an example, five FTU nodes are added based on the deployment location of the FTU, numbered 34 to 38. At the same time, the branches separated by the FTU are redefined based on the positive direction of the main power current. At this time, the topology contains a total of 38 nodes, 38 branches, and 5 tie lines. The nodes are represented by V = {v1, v2, ..., v 38} indicates that nodes 1 to 33 are TTU nodes, and nodes 34 to 38 are FTU nodes. Since there are two different types of nodes and the sum of the node and edge types is greater than 2, the distribution network topology at this time is heterogeneous.

[0062] Since the fault characteristics are the least obvious when a single-phase grounding fault occurs in a distribution network where the neutral point is not directly grounded, the present invention focuses on simulating three types of single-phase grounding faults at different locations. Figure 5 A simulation model was established for the topology of a power distribution network. A distribution network simulation model was constructed with the neutral point grounded via an arc suppression coil. Two distributed photovoltaic power sources were connected at nodes 18 and 29 to form an active distribution network. The system's base voltage was set to 12.66 kV, the base power to 20 MW, and a dynamic load model and PV-based distributed photovoltaic power sources were used. Each DG had an active power output of 1 MW and a reactive power output of 0.

[0063] The topologies used are shown in Table 1. Topologies numbered 1 to 6 are used to train the model to obtain fault data. Topologies numbered 7 to 10 are used as unknown topologies to verify the adaptability of the model under unknown topologies.

[0064] Table 1 Topology used

[0065] Topology Number Cut off branch roads Connecting branch 1 - - 2 S19 S33 3 S9 S34 4 S9, S15 S35, S36 5 S14, S28 S34, S37 6 S14, S20 S33, S36 7 S9 S36 8 S14, S28 S34, S37 9 S14, S20 S33, S36 10 S10, S19, S27 S34, S35, S36

[0066] S1.2. Obtain measurement data. For the intelligent distribution transformer terminal (TTU) node, the measurement data obtained includes three-phase voltage, three-phase current, zero-sequence voltage, and zero-sequence voltage. For the feeder terminal unit (FTU) node, the measurement data obtained includes three-phase voltage, three-phase current, zero-sequence voltage, zero-sequence voltage, and fault current direction. The three-phase voltage, three-phase current, zero-sequence voltage, and zero-sequence voltage are the amplitude data within the measurement period (5 minutes). The fault current direction data value is "1", "0", or "-1", indicating that the fault current direction is from the bus to the line, no fault current is detected, and the fault current direction is from the line to the bus, respectively. When obtaining the fault data set, the fault resistance is set to 0.1Ω and the initial fault phase angle is set to 0°. 100 short-circuit fault simulations are performed on each of the 32 branches except S8, S12, S18, S22, and S25. To better reflect actual conditions, single-phase grounding faults accounted for 60%, two-phase short circuits and two-phase short-circuit-to-ground faults each accounted for 15%, and three-phase short circuits and three-phase short-circuit-to-ground faults each accounted for 5%, resulting in a total of 3,200 sets of fault data. To ensure the completeness of the dataset, the dataset was augmented with normal operation data, bringing the total to 5,000 sets for each topology.

[0067] S2. Preprocess the measured electrical data and construct a fault data set;

[0068] S2.1. Preprocessing of measurement data. The main purpose of preprocessing of measurement data is to normalize the node information. Figure 2 The node set V in the 38}, the obtained set of original measurement data is X raw ={x raw1 ,x raw2 ,...,x raw38}.

[0069] First, the missing parts in the acquired measurement data should be filled with zero values, and the missing values ​​such as NaN and NULL in the data should be replaced with 0 to complete the missing parts.

[0070] Secondly, X raw ={x raw1 ,x raw 2,...,x raw38}, perform standard deviation normalization (Z-score) or linear normalization (Min-Max) to prevent dimensional differences from affecting model training. Because there are two types of nodes, TTU and FTU, and the data dimensions of the two types of nodes are different, the measurement data of different types of nodes needs to be normalized separately.

[0071] For a set of data X={x1,x2,...,x n}, the formula for standard deviation normalization is:

[0072]

[0073] in, is x i The new value after normalization, μ is the mean of data X σ is the standard deviation of the data X

[0074] The formula for linear normalization is:

[0075]

[0076] Among them, x p Represents each value in the data, Represents the new value after normalization.

[0077] S2.2. Construct a fault data set, specifically:

[0078] Get the node information set X; for a graph structure with n nodes, the node set V in the graph is V = {v1, v2, ..., v n}, the set of node information X is X={x1,x2,…,x n}, for each node's node information x i ,i∈[1,n], which is composed of node type x node , node number x ID , node feature x feat with node label x lable Together, they constitute x i =(x node ,x ID ,x feat ,x lable );

[0079] Get the edge information set E, for the edge information set E, E=(e1,e2,…,e m ), where m represents the number of edges, and the edge information e for each edge j ,j∈[1,m], which is composed of the starting node number, the target node number and the edge type, that is, e j =(v ID_start ,v ID_to ,e type ), where edge type e type There are three categories: represents the edge from FTU node to TTU node; Represents the edge from TTU node to FTU node; represents the edge from TTU node to TTU node;

[0080] Define heterogeneous graph path Φ and construct fault information label L.

[0081] The heterogeneous graph element path Φ is composed of multiple groups of edge types. According to the actual situation of the distribution network, two groups of edge types are used to construct possible element paths, as follows:

[0082] The specific path of the first meta-path Φ1 is: from FTU node to TTU node and then to TTU node:

[0083] The specific path of the second meta-path Φ2 is: from TTU node to FTU node and then to TTU node:

[0084] The specific path of the third element path Φ3 is: from TTU node to TTU node and then to FTU node:

[0085] The specific path of the fourth element path Φ4 is: from TTU node to TTU node and then to TTU node:

[0086] If a distribution network incorporates distributed generation (DGs), large capacity can lead to reverse power flow. In the event of a fault, DGs will also deliver current to the fault point. Therefore, in a distribution network, especially at the end nodes of a line, power transmission can be bidirectional: from the busbar to the branch, or vice versa. Therefore, when constructing the distribution network topology, an undirected graph is used. The meta-paths designed in Table 2 accurately represent the possible power transmission path relationships.

[0087] Table 2 Metapath Design

[0088]

[0089]

[0090] The fault information tag is composed of L=(L location ,L type ,L phase ), where the fault status label L location For fault location tasks, the fault status label is defined as follows: when a short circuit fault occurs on a line, the fault status label of the node not affected by the short circuit fault is "1", and the fault status label of the node affected by the short circuit fault is "0";

[0091] Fault category label L typeFor fault category classification tasks, the fault categories include no fault, single-phase grounding fault, two-phase interphase short circuit fault, two-phase short circuit grounding fault, three-phase short circuit fault, and three-phase short circuit grounding fault. The fault category labels are encoded using a one-hot encoding method to map the six fault categories into a six-dimensional binary vector. The specific encoding method is shown in Table 3.

[0092] Table 3 Fault category codes

[0093] Fault type <![CDATA[L type Corresponding one-hot encoding]]> No trouble [1,0,0,0,0,0] Single-phase ground fault [0,1,0,0,0,0] Two-phase short circuit fault [0,0,1,0,0,0] Two-phase short circuit ground fault [0,0,0,1,0,0] Three-phase short circuit fault [0,0,0,0,1,0] Three-phase short circuit ground fault [0,0,0,0,0,1]

[0094] Fault phase label L phase For fault phase classification tasks, the fault phases include no fault phase, A phase fault, B phase fault, C phase fault, AB two-phase fault, BC two-phase fault, CA two-phase fault, and ABC three-phase fault. The fault phase labels are mapped into 8-dimensional binary vectors using a one-hot encoding method. The specific encoding rules are shown in Table 4.

[0095] Table 4 Fault phase codes

[0096]

[0097]

[0098] S3. Construct a distribution network short-circuit fault diagnosis model based on heterogeneous graph attention network; the distribution network short-circuit fault diagnosis model includes a fault location model, a fault category and a fault phase classification model, and the output of each model is used as the complete fault diagnosis result.

[0099] S3.1. Constructing a fault location model

[0100] The fault localization model uses a two-layer heterogeneous graph attention network (HGAL) as a feature encoder to capture features of different fault locations. A multi-layer perceptron (MLP) is used as the output layer, and a sigmoid activation function is used as the final output layer to output the probability of each node being in a faulty state. Furthermore, a skip connection layer is introduced between the model input and the second HGAL layer to improve the model's gradient propagation efficiency, mitigate degradation issues in deep networks, and enhance feature reuse, thereby improving the model's representation accuracy of fault location features and training stability.

[0101] The specific structure of the fault localization model based on heterogeneous graph attention network is as follows Figure 3 shown.

[0102] In the model, each HAGL layer consists of two levels of attention: node-level attention and semantic-level attention.

[0103] (1) Node-level attention

[0104] In heterogeneous graphs, the core idea of ​​node attention is to dynamically assign different attention weights based on the importance of the node’s neighbors in different meta-paths. i ,v j ), node correlation and node attention coefficient Can be expressed as:

[0105]

[0106] Among them, x' i and x' j They are node v i 、v j Information x i 、x j After the dimension transformation, it is mapped to the new features of the same dimension, N(v i ) Φ v is the metapath Φ i The number of neighbor nodes, is the learnable node attention weight matrix.

[0107] Similarly, through the aggregation of attention mechanism, based on the meta-path Φ, node v i The aggregation of its neighbor features can be expressed as:

[0108]

[0109] in, is node v i The new node-level features learned through the meta-path Φ. Since a single-group attention mechanism cannot fully learn the complex dependencies between node features, a multi-head attention mechanism can effectively solve this problem. The node attention mechanism is the middle part of the heterogeneous graph attention mechanism, so a spliced ​​multi-head attention mechanism is used. For an attention mechanism with M attention heads, it can be updated as follows:

[0110]

[0111] Among them, “||” represents the concatenation operation and σ(·) represents the activation function.

[0112] For a given set of meta-paths Φ={Φ0,Φ1,…,Φ P}, after the original features of the node are input into the node-level attention mechanism, the node feature set Z under the specific semantics of the P group can be obtained = {Z Φ0 ,Z Φ1 ,…,Z ΦP}.

[0113] For node features n is the total number of nodes, which represents the new node feature representation obtained by aggregating the neighbor information on the meta-path Φ of all nodes in the heterogeneous graph.

[0114] (2) Semantic-level attention

[0115] After obtaining the feature representation of the node under different meta-paths through node-level attention, the semantic attention mechanism takes it as input and calculates a semantic-level attention weight for each meta-path to represent the importance of the meta-path to the target node. The final node feature is obtained by weighted aggregation of the features of different meta-paths.

[0116] First, before calculating the attention coefficients of different meta-paths on the target node, the node features under each meta-path Φ should be nonlinearly mapped to a new feature space to enhance its expressive power, namely:

[0117]

[0118] in, After nonlinear transformation That is, the node v after nonlinear transformation i The new node-level feature representation learned by the meta-path set Φ, W and b are the learnable weight matrix and bias vector for nonlinear transformation, respectively.

[0119] Secondly, use With w sem The similarity between the two is measured by the dot product operation result. In order to eliminate the influence of the number of nodes, the average of all nodes should be taken, so that each meta-path Φ p,p∈[1,P] Relevance score Right now:

[0120] In order to eliminate the influence of different numbers of nodes, the average of all nodes should be taken. Then for the meta-path set Φ={Φ0,Φ1,…,Φ P} a meta-path Φ p,p∈[1,P] , its correlation score It can be expressed as:

[0121]

[0122] Among them, w sem is a learnable semantic attention weight vector, Represents the node v after nonlinear transformation i Through the meta-path Φ p The obtained feature representation.

[0123] Similarly, to facilitate the comparison of the importance of meta-paths to nodes, the importance scores of meta-paths are normalized and the weight of each meta-path is calculated using the softmax function:

[0124]

[0125] Finally, the semantic-level node features obtained by weighted aggregation of different meta-paths are:

[0126]

[0127] Among them, Z i 'For node v i Feature information x i The new node features are output after passing through the HGAL layer.

[0128] The process of node attention mechanism and semantic attention mechanism aggregating features is as follows Figure 4 shown.

[0129] S3.2. Construct a classification model for fault categories and fault phases;

[0130] The fault category and fault phase classification model uses a two-layer heterogeneous graph attention network (HGAL) as a feature encoder to capture the characteristics of different fault categories and fault distinctions. A global average pooling layer is introduced before the output layer to average all node features and aggregate them into features for the entire graph. A fully connected layer (FCL) serves as the output layer, using Softmax as the activation function to predict the output as the probabilities of different categories. Similarly, a skip connection layer is introduced between the model input and the second HGAL layer. Through dimensional mapping, low-level features are directly transferred to higher levels to avoid vanishing or exploding gradients during training.

[0131] The specific structure of the fault category and fault phase classification model based on heterogeneous graph attention network is as follows: Figure 5 shown.

[0132] S4. Define the training hyperparameters and loss function of the distribution network short-circuit fault diagnosis model, divide the fault data set into a training set, a validation set, and a test set, use the training set and the test set to train the distribution network short-circuit fault diagnosis model, fine-tune the parameters during training, and use the validation set to verify the performance of the trained distribution network short-circuit fault diagnosis model to obtain the optimal distribution network short-circuit fault diagnosis model;

[0133] S4.1. Define model hyperparameters

[0134] For distribution network short-circuit faults, the fault location model uses two HGAL layers with 64 and 128 neurons, respectively. The MLP contains two hidden layers with 128 and 64 neurons, respectively. The Sigmoid dimension is 33. In the fault category and fault phase classification models, the two HGAL layers use 32 and 64 neurons, respectively. The FCL layer uses 32 neurons. The Softmax dimension is 6 in the fault category classification model and 8 in the fault phase classification model. The specific hyperparameter settings are shown in the following table:

[0135] Table 5 Model hyperparameters

[0136]

[0137] S4.2. Define the model loss function.

[0138] Since in actual situations, failures are a minority, and the data set obtained contains the majority of normal operating state data, it is an extremely unbalanced data set. If a general cross-entropy loss function is used for training, the model will tend to learn the features of the majority class and ignore the minority class, resulting in poor classification performance for fault samples and difficulty in capturing key abnormal patterns. It is necessary to introduce data balancing technologies such as weighted cross-entropy loss, focal loss, or oversampling and undersampling to alleviate the class imbalance problem, thereby improving the model's recognition ability for the minority class and improving the overall classification effect. In this invention, a focal loss function with an adaptive modulation factor is designed as the loss function of the model. The specific expression is as follows:

[0139]

[0140] Where, Loss FL is the focal loss, is the adaptive modulation factor, p i,k is the probability that the i-th sample belongs to category k; N is the total number of samples, K is the total number of categories, y i,k is a label used to mark whether the i-th sample belongs to category k. If it does, y i,k =1, otherwise y i,k =0;γ init is the initial modulation factor set; N k_samples is the total number of samples in the kth category; i iteration The number of iterations during model training.

[0141] By setting the adaptive modulation It can be achieved that in the early stage of training when the number of iterations is small, the modulation factors of samples of different numbers are large, thereby accelerating the model's learning of complex features; as the number of iterations increases, in the later stage of training, the modulation factors of samples of different categories gradually decrease, so that the model can balance the contribution of all samples to model learning, avoid over-focusing on difficult samples and neglecting easy samples, and ensure that the overall performance of the model is improved. At the same time, to address the problem of class imbalance, the adaptive modulation factor can make the modulation factor of categories with fewer samples higher, so that the model can better learn the features in these categories; for categories with more samples, the value of the modulation factor can be appropriately reduced to reduce its impact on model training. In this way, the focal loss function with an adaptive modulation factor can effectively solve the problem of imbalance in the number of samples between categories while maintaining the focal loss's focus on difficult-to-classify samples, thereby improving the overall performance of the model in imbalanced data scenarios.

[0142] The changes in the modulation factor under different sample numbers are as follows: Figure 6 shown.

[0143] S4.3. Training distribution network fault diagnosis model

[0144] After determining the model hyperparameters and loss function, the dataset was divided into a training set, a test set, and a validation set in a ratio of 8:1:1. The training and validation sets were used for model training and training performance evaluation, while the validation set was not used in training. It was used after model training to verify the model's accuracy and generalization ability on unlearned data.

[0145] During the parameter update process, the Adam optimizer is used to perform adaptive adjustment of the parameters to achieve a more efficient and stable training process. When the loss function of the model gradually flattens and approaches 0, the model training is completed.

[0146] S5. When a distribution network fault occurs, the real-time measurement data is preprocessed and input into the optimal distribution network short-circuit fault diagnosis model to obtain the prediction results of the node fault state, fault type and fault phase.

[0147] Example 2

[0148] The present invention proposes an electronic device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the method for diagnosing short-circuit faults in a distribution network based on a heterogeneous graph attention network as described in Example 1 is implemented.

[0149] Example 3

[0150] The present invention proposes a computer-readable storage medium storing a computer program, wherein the computer program enables a computer to execute the distribution network short-circuit fault diagnosis method based on a heterogeneous graph attention network as described in Example 1.

[0151] The effectiveness of the present invention is verified by examples below.

[0152] (1) Comparison of fault location performance

[0153] To validate the effectiveness of the HGAT-based fault localization model, we trained models based on GCN and GAT simultaneously for comparison with the proposed model. Notably, since neither model is suitable for heterogeneous graph environments, we only used TTU nodes to construct the corresponding homogeneous graph during training and validation, ignoring the feature data of the FTU nodes in the dataset. The evaluation metrics for each model on the validation set are shown in Table 6:

[0154] Table 6 Accuracy of different models

[0155] Algorithms Accuracy / % Precision / % Recall / % F1-score / % CNN 85.90 83.56 84.62 84.09 GCN 90.98 90.13 90.82 90.47 GAT 94.08 92.43 93.14 92.78 HGAT-FL 98.52 97.29 96.91 97.10 HGAT-ATFL 99.31 99.12 98.81 98.96

[0156] As shown in Table 6, the fault localization model based on HGAT outperforms the GCN based on homogeneous graphs and the GAT based on graph attention networks in all evaluation metrics. This is due to HGAT's ability to fully learn the features of different dimensions between FTUs and TTUs and integrate fault current direction information for more accurate fault localization. In contrast, GCN and GAT, designed only for homogeneous graphs, can only process features of homogeneous graphs and uniform dimensions, making limited use of information. Therefore, their performance is relatively low under the same conditions.

[0157] When comparing the model performance of HGAT and the HGAT-ATFL proposed in this invention, it can be observed that HGAT-ATFL outperforms HGAT in all performance indicators, which shows that the adaptive focus loss can effectively deal with the problem of class imbalance, especially when facing difficult classification and a small number of samples. This will further improve the classification performance of the model and make the model show greater advantages in overall accuracy and robustness.

[0158] (2) Fault identification effect analysis and feature visualization

[0159] The fault identification task includes the identification of fault categories and fault differences. The identification results of different models are shown in the following table:

[0160] Table 7 Accuracy of different models

[0161]

[0162] As shown in Table 7, all models outperformed the fault location task under the same circumstances for fault identification. This is because fault identification, as a whole-graph classification task, has fewer categories than the fault location task, making the model learning easier and improving performance. The model based on the heterogeneous graph attention network has stronger semantic expression capabilities and can learn the rich topological relationships and fault feature information in the distribution network, thereby more accurately identifying fault categories and fault phases. The F1-score for the two identification tasks reached 0.9968 and 0.9934, respectively, which is more than 1% higher than the GCN and GAT identification models using feature data of the same dimensionality, achieving superior fault identification results. This result demonstrates that the heterogeneous graph attention network has a strong advantage in handling complex fault scenarios in distribution networks and can effectively improve the accuracy and reliability of fault detection.

[0163] At the same time, the present invention uses the t-SNE method to perform two-dimensional visualization of the node features before the output layer of the trained HGAT-based fault category identification and fault phase identification model, as shown in Figure 7. It can be seen that no matter whether it is fault category or fault phase identification, after the dimensionality reduction of the node features, each category can still form a compact cluster. This shows that the heterogeneous graph attention network can effectively extract features of different categories and separate them in the feature space through continuous learning of fault features, and has a strong feature learning ability. At the same time, there are very few outliers in other category clusters. This may be because the features of some categories may have a high similarity in the original data, which makes the model's feature extraction of these data insufficient. Despite this, overall, the application of the heterogeneous graph attention network in fault category and fault phase identification tasks still shows high robustness and discrimination ability.

[0164] (3) Model’s adaptability to topology reconstruction

[0165] In order to further clarify the performance advantage of the HGAT-based distribution network fault location model under the condition of distribution network topology reconstruction, the unknown topology structures with topology numbers 7 to 10 in Table 1 are used to simulate short-circuit faults, and the corresponding fault data test set under the unknown topology is constructed. Without retraining the model, the fault location accuracy of the HAGT, GAT and GCN models is tested, and the results are as follows: Figure 8 shown.

[0166] Depend on Figure 8As can be seen, the F1-score values ​​of each model exhibit varying degrees of fluctuation when faced with different unknown topologies. For situations with minimal topological changes, HGAT-ATFL consistently maintains an F1-score above 0.85, and even maintains performance above 0.8 for more complex topological changes. In contrast, the GCN and GAT models experience a significant decline in localization performance when faced with unknown topologies, with the performance loss being particularly pronounced under complex topological changes. HGAT-ATFL's superior performance stems from its unique heterogeneous nature. Through its heterogeneous graph attention mechanism, the model accurately captures the complex relationships between heterogeneous nodes and employs a hierarchical attention mechanism to dynamically adjust the importance of nodes and meta-paths. Compared to the GCN model, it is more effective in handling heterogeneous networks. Compared to GAT, it not only focuses on node features but also deeply integrates heterogeneous information, resulting in greater stability in the face of topological changes. This design enables HGAT-ATFL to demonstrate superior adaptability and generalization in complex and changing topological environments, providing a new solution to the problem of fault localization effectiveness in the face of topological reconfiguration.

[0167] In the embodiments disclosed herein, computer storage media can be tangible media that can contain or store programs for use by or in conjunction with an instruction execution system, device, or apparatus. Computer storage media can include, but are not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices, or equipment, or any suitable combination of the foregoing. More specific examples of computer storage media can include electrical connections based on one or more lines, portable computer disks, hard disks, random access memories (RAM), read-only memories (ROM), erasable programmable read-only memories (EPROM or flash memory), optical fibers, portable compact disk read-only memories (CD-ROMs), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0168] Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed in this application can be implemented in electronic hardware or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0169] The above are merely preferred embodiments of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions based on the principles of the present invention are within the scope of protection of the present invention. It should be noted that for those skilled in the art, various improvements and modifications that do not depart from the principles of the present invention should be considered within the scope of protection of the present invention.

Claims

1. A distribution network short circuit fault diagnosis method based on heterogeneous graph attention network, characterized in that: The following steps are involved: Establish a heterogeneous distribution network topology structure and measure the electrical data of each node in the distribution network topology; Preprocess the measured electrical data and construct a fault dataset; Construct a distribution network short circuit fault diagnosis model based on heterogeneous graph attention network; Define the training hyperparameters and loss function of the distribution network short-circuit fault diagnosis model. Divide the fault dataset into training, validation, and test sets. Use the training and test sets to train the distribution network short-circuit fault diagnosis model. Fine-tune the parameters during training. Use the validation set to verify the performance of the trained distribution network short-circuit fault diagnosis model and obtain the optimal distribution network short-circuit fault diagnosis model. When a distribution network fault occurs, the real-time measurement data is preprocessed and input into the optimal distribution network short-circuit fault diagnosis model to obtain the prediction results of the node fault state, fault type and fault phase.

2. The distribution network short circuit fault diagnosis method based on heterogeneous graph attention network according to claim 1 is characterized in that: The specific structure of establishing a heterogeneous distribution network topology is as follows: The two types of measuring devices in the distribution network, intelligent distribution transformer terminals (TTUs) and feeder terminal units (FTUs), are used as nodes in the distribution network topology. The power transmission relationships between the measuring devices are used as edges in the distribution network topology. Based on the distance from the substation, the nodes are numbered from far to near, from the main line to the branch, and first the TTU nodes and then the FTU nodes to construct a distribution network topology diagram. The distribution network topology diagram satisfies the requirement that the sum of the node and edge types is greater than 2.

3. The distribution network short circuit fault diagnosis method based on heterogeneous graph attention network according to claim 1 is characterized in that: The electrical data of each node in the distribution network topology diagram is specifically as follows: For the intelligent distribution transformer terminal TTU node, the measurement data obtained include three-phase voltage, three-phase current, zero-sequence voltage and zero-sequence voltage; for the feeder terminal unit FTU node, the measurement data obtained include three-phase voltage, three-phase current, zero-sequence voltage, zero-sequence voltage and fault current direction; among them, the three-phase voltage, three-phase current, zero-sequence voltage and zero-sequence voltage are the amplitude data within the measurement time period, and the value of the fault current direction data is "1", "0" or "-1", which respectively indicate that the fault current direction is from the bus to the line, no fault current is detected, and the fault current direction is from the line to the bus.

4. The method for diagnosing short-circuit faults in distribution networks based on heterogeneous graph attention networks according to claim 1, wherein: The preprocessing of the measured electrical data is specifically as follows: The missing parts of the measured electrical data are filled with zero values ​​to fill the missing parts, and Z-score or Min-Max normalization is performed on them.

5. The distribution network short circuit fault diagnosis method based on heterogeneous graph attention network according to claim 1 is characterized in that: The construction of the fault data set is specifically as follows: Get the node information set X; for a graph structure with n nodes, the node set V in the graph is V = {v1, v2, ..., v n }, the set of node information X is X={x1,x2,...,x n }, for each node's node information x i ,i∈[1,n], which is composed of node type x node , node number x ID , node feature x feat with node label x lable Together, they constitute x i =(x node ,x ID ,x feat ,x lable ); Get the edge information set E, for the edge information set E, E=(e1,e2,...,e m ), where m represents the number of edges, and the edge information e for each edge j ,j∈[1,m], which is composed of the starting node number, the target node number and the edge type, that is, e j =(v ID_start ,v ID_to ,e type ), where edge type e type There are three categories: represents the edge from FTU node to TTU node; Represents the edge from TTU node to FTU node; represents the edge from TTU node to TTU node; Define a heterogeneous graph element path Φ and construct a fault information label L. The heterogeneous graph element path Φ is composed of multiple groups of edge types. According to the actual situation of the distribution network, a possible element path is constructed by splicing two groups of edge types, as follows: The specific path of the first meta-path Φ1 is: from FTU node to TTU node and then to TTU node: The specific path of the second meta-path Φ2 is: from the TTU node to the FTU node and then to the TTU node: The specific path of the third element path Φ3 is: from TTU node to TTU node and then to FTU node: The specific path of the fourth element path Φ4 is: from TTU node to TTU node and then to TTU node: The fault information tag is composed of L=(L location ,L type ,L phase ), where the fault status label L location For fault location tasks, the fault status label is defined as follows: when a short circuit fault occurs on a line, the fault status label of the node not affected by the short circuit fault is "1", and the fault status label of the node affected by the short circuit fault is "0"; Fault category label L type It is used for fault classification tasks. The fault categories include no fault, single-phase grounding fault, two-phase interphase short circuit fault, two-phase short circuit grounding fault, three-phase short circuit fault, and three-phase short circuit grounding fault. The fault category labels are mapped into 6-dimensional binary vectors using one-hot encoding. Fault phase label L phase It is used for fault phase classification tasks. Fault phases include no fault phase, A phase fault, B phase fault, C phase fault, AB two-phase fault, BC two-phase fault, CA two-phase fault, and ABC three-phase fault. The fault phase labels are encoded using a one-hot encoding method, mapping the eight fault phases into an 8-dimensional binary vector.

6. The distribution network short circuit fault diagnosis method based on heterogeneous graph attention network according to claim 1, characterized in that: The construction of the distribution network short circuit fault diagnosis model based on heterogeneous graph attention network is specifically as follows: The distribution network short-circuit fault diagnosis model includes a fault location model and a fault type and fault phase classification model; The fault localization model uses two layers of heterogeneous graph attention networks (HGALs) as feature encoders to capture features of different fault locations. A multi-layer perceptron (MLP) is used as the output layer, and Sigmoid is used as the activation function of the final output layer. The output layer outputs the probability that each node is in a fault state. A skip connection layer is introduced between the input of the fault localization model and the second HGAL layer to directly transfer low-level features to high-level features through dimensional mapping. The fault category and fault phase classification model uses a two-layer heterogeneous graph attention network (HGAL) as a feature encoder to capture the features of different fault categories and fault phases. A global average pooling layer is introduced before the output layer to average all node features and aggregate them into features for the entire graph. A fully connected layer (FCL) is used as the output layer, and Softmax is used as the activation function of the output layer. The output layer outputs the probability distribution of the fault category and fault phase for different categories. A skip connection layer is introduced between the input of the fault category and fault phase classification model and the second HGAL layer to directly transfer low-level features to high-level features through dimensional mapping.

7. The distribution network short circuit fault diagnosis method based on heterogeneous graph attention network according to claim 1 is characterized in that: The loss function of the distribution network short-circuit fault diagnosis model adopts a focal loss function with an adaptive modulation factor. The specific expression is as follows: Where, Loss FL is the focal loss, is the adaptive modulation factor, p i,k is the probability that the i-th sample belongs to category k; N is the total number of samples, K is the total number of categories, y i,k is a label used to mark whether the i-th sample belongs to category k. If it does, y i,k =1, otherwise y i,k =0;γ init is the initial modulation factor set; N k_samples is the total number of samples in the kth category; i iteration The number of iterations during model training; During the parameter updating process, the Adam optimizer is used to perform adaptive adjustment of the parameters. The training ends when the loss function value of the distribution network short-circuit fault diagnosis model converges.

8. The distribution network short circuit fault diagnosis method based on heterogeneous graph attention network according to claim 6, characterized in that: Each heterogeneous graph attention network layer HGAL in the fault localization model includes two levels of attention: node-level attention and semantic-level attention. The node-level attention dynamically assigns different attention weights based on the importance of the node's neighbors in different meta-paths. After obtaining the feature representation of the node under different meta-paths through node-level attention, semantic-level attention takes it as input and calculates a semantic-level attention weight for each meta-path to indicate the importance of the meta-path to the target node. The final node feature is obtained by weighted aggregation of the features of different meta-paths.

9. An electronic device, characterized in that: include: A memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the method for diagnosing short-circuit faults in a distribution network based on a heterogeneous graph attention network as described in any one of claims 1 to 8 is implemented.

10. A computer-readable storage medium storing a computer program, characterized in that: The computer program enables a computer to execute the distribution network short circuit fault diagnosis method based on heterogeneous graph attention network as described in any one of claims 1 to 8.

Images