Power grid fault location method, system, device and product based on graph neural network
By using a graph neural network-based power grid fault location method, a graph model is generated and a spatiotemporal graph convolutional network and a graph neural network are trained. The features of power grid nodes are extracted and mapped to edge features, which solves the problem of low accuracy of traditional methods in complex power grids and achieves efficient fault location.
Patent Information
- Application Number
- CN202510499660.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2025-04-21
- Publication Date
- 2026-02-13
- Estimated Expiration
- 2045-04-21
AI Technical Summary
Traditional fault location methods have low accuracy and low fault tolerance in complex power grids, and are difficult to adapt to changes in power grid topology, resulting in low fault location efficiency and inability to meet real-time requirements.
A graph neural network-based approach is adopted. By generating a graph model, a spatiotemporal graph convolutional network and a graph neural network are trained to extract the features of power grid nodes. Then, a one-dimensional convolutional neural network is used to map the features to the edge features to calculate the probability value of the fault section.
It significantly improves the accuracy of fault location, has strong robustness, can cope with topology changes and high-resistance faults, and adapts to scenarios with incomplete data.
Smart Images
Figure CN120408196B_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The application belongs to the technical field of fault location, and particularly relates to a power grid fault location method, system, device and product based on a graph neural network. BACKGROUND
[0002] In the complex network architecture of the power system, the distribution network is a key link directly facing users, and its safe and stable operation is directly related to the normal order of social production and life. Quickly and accurately locating the fault section is not only an important prerequisite for ensuring power supply reliability, but also a core measure to improve the emergency response capability of the power grid and reduce power loss. However, traditional fault location methods often face many challenges. On the one hand, these methods mostly rely on human experience or simple threshold judgment, and have limited recognition ability for complex and variable fault characteristics, low fault tolerance, and are prone to misjudgment or omission in the presence of noise interference or signal distortion, resulting in a significant reduction in the accuracy of fault location. On the other hand, traditional methods are inefficient in processing large-scale data, and the fault location process is time-consuming, which is difficult to meet the high real-time requirements of modern power systems.
[0003] With the rapid development of artificial intelligence technology, deep learning algorithms have shown great potential in fault diagnosis. However, most deep learning models still have significant limitations when applied to distribution network fault location. Especially in the face of the actual situation of frequent changes in power grid topology, these models often lack sufficient generalization ability. Since deep learning models are highly dependent on the distribution characteristics of training data, once the power grid structure changes, such as the addition of new lines or equipment modification, the original model may fail to adapt to the new data characteristics, resulting in performance degradation or even failure.
[0004] Therefore, based on the foregoing deficiencies, how to provide a power grid fault location method, system, device and product based on a graph neural network that can improve the accuracy of distribution network fault location has become a problem to be solved. SUMMARY
[0005] The purpose of the present application is to provide a power grid fault location method, system, device and product based on a graph neural network to solve the problems of serious eddy current or residual magnetism influence and long scanning time existing in the prior art.
[0006] In order to achieve the above-mentioned purpose, the technical scheme adopted by the present application is as follows:
[0007] In a first aspect, the present application provides a power grid fault location method based on a graph neural network, which comprises:
[0008] An actual power grid line is acquired, and a graph model is generated according to the actual power grid line, wherein the graph model comprises a node set and an edge line set, the node set is used to represent various electrical operating devices, and the edge line set is used to represent power line sections;
[0009] Telemetry data and telecontrol data corresponding to each electrical operating device are respectively acquired, and the telemetry data and the telecontrol data corresponding to each electrical operating device are respectively stored in each node of the node set, wherein the telemetry data comprises voltage signals and current time sequence signals of the corresponding electrical operating device, and the telecontrol data comprises discrete action information of the corresponding electrical operating device;
[0010] Based on the node set, a graph neural network model and a spatio-temporal graph convolution network model based on attention are trained, and the node set is processed according to the graph neural network model and the spatio-temporal graph convolution network model based on attention to extract power grid node features;
[0011] The power grid node features are input into a one-dimensional convolutional neural network, and a one-dimensional convolution kernel is used for node coverage to map the power grid node features to power grid edge line features;
[0012] The power grid edge line features are dimensionally reduced, and probability values of each edge line in the edge line set being a fault section are output, and a corresponding edge line with the largest probability value is selected as a power grid fault positioning result.
[0013] In a possible design, based on the node set, the graph neural network model and the spatio-temporal graph convolution network model based on attention are trained, comprising:
[0014] Telemetry data of each node in the node set is acquired;
[0015] A first training set is formed by the telemetry data of each node;
[0016] Based on the first training set, the graph neural network model based on attention is trained;
[0017] Telecontrol data of each node in the node set is acquired;
[0018] A second training set is formed by the telecontrol data of each node;
[0019] Based on the second training set, the graph neural network model is trained.
[0020] In a possible design, according to the graph neural network model and the spatio-temporal graph convolution network model based on attention, the node set is processed to extract power grid node features, comprising:
[0021] projecting the node set to a time attention space based on the spatio-temporal graph convolution network model, calculating time attention weights of each node in the node set respectively according to the time attention space projection of the node set, and obtaining a time attention matrix based on the time attention weights of each node;
[0022] projecting the node set to a spatial attention space based on the spatio-temporal graph convolution network model, calculating spatial attention weights of each node in the node set respectively according to the spatial attention space projection of the node set, and obtaining a spatial attention matrix based on the spatial attention weights of each node;
[0023] processing the spatial attention matrix based on the graph convolution method to extract spatial features of each node in the node set;
[0024] extracting time features of each node in the node set based on the time attention matrix;
[0025] obtaining spatio-temporal fusion features of each node in the node set based on the spatial features of each node and the time features of each node in the node set.
[0026] In a possible design, after obtaining the spatio-temporal fusion features of each node in the node set based on the spatial features of each node and the time features of each node, the method further includes:
[0027] constructing a multi-head graph attention mechanism for the node set based on the graph neural network model;
[0028] calculating graph attention weights of each node in the node set by using the multi-head graph attention mechanism;
[0029] processing each node in the node set according to the graph attention weights of each node in the node set, to calculate graph attention features by each attention head in the multi-head graph attention mechanism;
[0030] concatenating the graph attention features calculated by each attention head with the spatio-temporal fusion features of each node in the node set, to obtain power grid node features.
[0031] In a possible design, projecting the node set to a time attention space based on the spatio-temporal graph convolution network model, calculating time attention weights of each node in the node set respectively according to the time attention space projection of the node set, and obtaining a time attention matrix based on the time attention weights of each node, includes:
[0032] Using a learnable temporal linear transformation matrix, the telemetry data of each node in the node set are linearly processed to project the node set into a temporal attention space.
[0033] The temporal similarity matrix between nodes is calculated based on the temporal attention space projection of the node set.
[0034] A time bias term matrix is added to the time similarity matrix between the nodes, and a time attention weight matrix is generated using an activation function;
[0035] Using a learnable temporal weight matrix, the temporal attention weight matrix is linearly processed to calculate the temporal attention weight of each node in the node set, and the temporal attention matrix is obtained based on the temporal attention weight of each node.
[0036] Accordingly, based on the spatiotemporal graph convolutional network model, the node set is projected onto the spatial attention space. Based on the projection of the node set into the spatial attention space, the spatial attention weights of each node in the node set are calculated, and a spatial attention matrix is obtained based on the spatial attention weights of each node. This matrix includes:
[0037] Using the time attention matrix, each node in the node set is weighted to obtain a time-weighted feature matrix;
[0038] Using a learnable spatial linear transformation matrix, each node in the node set is linearly processed to project the node set onto the spatial attention space;
[0039] Based on the spatial attention spatial projection of the node set, the spatial similarity matrix between each node is calculated;
[0040] A spatial bias term matrix is added to the spatial similarity matrix between the nodes, and a spatial attention weight matrix is generated using an activation function.
[0041] The spatial attention weight matrix is linearly processed using a learnable spatial weight matrix to calculate the spatial attention weight of each node in the node set, and the spatial attention matrix is obtained based on the spatial attention weight of each node.
[0042] In one possible design, the spatial attention matrix is processed based on a graph convolution method to extract the spatial features of each node in the node set, including:
[0043] The spatial attention matrix is normalized to obtain a normalized spatial attention matrix;
[0044] The Chebyshev polynomial is used to perform graph convolution processing on the normalized spatial attention matrix to extract the spatial features of each node in the node set.
[0045] Correspondingly, based on the time attention matrix, the time features of each node in the node set are extracted, and the time features of each node in the node set are extracted.
[0046] The time attention matrix is processed by two-dimensional convolution to extract the time features of each node in the node set.
[0047] In one possible design, the power grid node features are input into a one-dimensional convolutional neural network, and a one-dimensional convolution kernel is used for node coverage to map the power grid node features to power grid edge features, including:
[0048] Based on the graph model, the power grid node features are topologically arranged to form a feature input matrix.
[0049] The feature input matrix is subjected to convolution operation.
[0050] The convolution kernel is used to slide and cover each adjacent node in the feature input matrix to extract the power grid edge features corresponding to the power grid node features.
[0051] In a second aspect, the present application provides a power grid fault location system based on a graph neural network, comprising:
[0052] A data acquisition unit is configured to acquire an actual power grid line and generate a graph model based on the actual power grid line, wherein the graph model includes a node set and an edge set, the node set is configured to represent each electrical operating device, and the edge set is configured to represent a power line section.
[0053] A data storage unit is configured to acquire telemetry data and telesignaling data corresponding to each electrical operating device respectively, and store the telemetry data and telesignaling data corresponding to each electrical operating device in each node of the node set respectively, wherein the telemetry data includes voltage signals and current time series signals of the corresponding electrical operating device, and the telesignaling data includes discrete action information of the corresponding electrical operating device.
[0054] A feature extraction unit is configured to train an attention-based spatio-temporal graph convolution network model and a graph neural network model based on the node set, and process the node set based on the attention-based spatio-temporal graph convolution network model and the graph neural network model to extract power grid node features, and input the power grid node features into a one-dimensional convolutional neural network and use a one-dimensional convolution kernel for node coverage to map the power grid node features to power grid edge features.
[0055] A fault positioning unit is configured to reduce dimensionality of power grid edge line features and output probability values of each edge line in the edge line set being a fault section, and select a corresponding edge line with the largest probability value as a power grid fault positioning result.
[0056] In a third aspect, the present application provides an electronic device, comprising a memory, a processor and a transceiver connected in sequence, wherein the memory is configured to store a computer program, the transceiver is configured to transceive messages, and the processor is configured to read the computer program and execute the power grid fault positioning method based on the graph neural network as in the first aspect or any possible design of the first aspect.
[0057] In a fourth aspect, the present application provides a computer program product comprising instructions which, when executed on a computer, cause the computer to execute the power grid fault positioning method based on the graph neural network as in the first aspect or any possible design of the first aspect.
[0058] Beneficial effects: The present application provides a power grid fault positioning method, system, device and product based on a graph neural network, which belongs to the technical field of fault positioning. Firstly, an actual power grid line is obtained, and a graph model is generated according to the actual power grid line. Then, the telemetry data and the telesignaling data corresponding to each electrical operating device are obtained respectively, and the telemetry data and the telesignaling data corresponding to each electrical operating device are stored in each node of the node set respectively. Based on the node set, a time-space graph convolution network model based on attention and a graph neural network model are trained, and the node set is processed according to the time-space graph convolution network model based on attention and the graph neural network model to extract power grid node features. Then, the power grid node features are input into a one-dimensional convolutional neural network, and node coverage is performed using a one-dimensional convolution kernel to map the power grid node features to power grid edge line features. Finally, the power grid edge line features are reduced in dimensionality, and probability values of each edge line in the edge line set being a fault section are output, and a corresponding edge line with the largest probability value is selected as a power grid fault positioning result. The scheme provided by the present application not only has strong robustness to topology changes, high-resistance faults and incomplete data scenarios, but also significantly improves the accuracy of power distribution network fault positioning. BRIEF DESCRIPTION OF DRAWINGS
[0059] Figure 1 A step flowchart of the power grid fault positioning method based on the graph neural network provided by the embodiments of the present application. DETAILED DESCRIPTION
[0060] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the present application will be briefly introduced below in combination with the drawings and the description of the embodiments or the prior art. Obviously, the following description of the drawings is only some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor on the basis of these drawings. It should be noted that the description of these embodiments is used to help understand the present application, but does not constitute a limitation on the present application.
[0061] It should be understood that although the terms first, second, etc. may be used herein to describe various elements, these elements should not be limited by these terms. These terms are only used to distinguish one element from another. For example, a first element can be called a second element, and similarly, a second element can be called a first element without departing from the scope of the example embodiments of the present application.
[0062] It should be understood that for the term "and / or" that may appear herein, it is only a description of the association relationship of the associated objects, which means that there can be three relationships, for example, A and / or B, which means that there are three cases of A alone, B alone, and A and B together; for the term " / and" that may appear herein, it is another description of the association relationship of another associated object, which means that there can be two relationships, for example, A / and B, which means that there are two cases of A alone and A and B together; in addition, for the character " / " that may appear herein, it generally means that the associated objects before and after are an "or" relationship.
[0063] Embodiments:
[0064] Referring to Figure 1 As shown in the figure, the power grid fault positioning method based on the graph neural network provided by the embodiment can be but is not limited to the following steps S100-S500:
[0065] S100. Obtain an actual power grid line, and generate a graph model according to the actual power grid line, wherein the graph model includes a node set and an edge line set, the node set is used to represent various electrical operating devices, and the edge line set is used to represent power line sections;
[0066] S200. Obtain the telemetry data and the telesign data corresponding to each electrical operating device respectively, and store the telemetry data and the telesign data corresponding to each electrical operating device into each node of the node set respectively, wherein the telemetry data includes voltage signals and current time sequence signals of the corresponding electrical operating device, and the telesign data includes discrete action information of the corresponding electrical operating device;
[0067] S300. Based on the node set, a spatio-temporal graph convolution network model based on attention and a graph neural network model are trained, and the node set is processed according to the spatio-temporal graph convolution network model based on attention and the graph neural network model to extract power grid node features.
[0068] In a possible implementation, in step S300, based on the node set, the spatio-temporal graph convolution network model based on attention and the graph neural network model are trained, including:
[0069] Obtaining telemetry data of each node in the node set;
[0070] Forming a first training set with the telemetry data of each node;
[0071] Based on the first training set, the spatio-temporal graph convolution network model based on attention is trained;
[0072] Obtaining telesign data of each node in the node set;
[0073] Forming a second training set with the telesign data of each node;
[0074] Based on the second training set, the graph neural network model is trained.
[0075] In a possible implementation, in step S300, according to the spatio-temporal graph convolution network model based on attention and the graph neural network model, the node set is processed to extract power grid node features, including:
[0076] Based on the spatio-temporal graph convolution network model, the node set is projected into a time attention space, the time attention weight of each node in the node set is calculated according to the time attention space projection of the node set, and the time attention matrix is obtained based on the time attention weight of each node;
[0077] Based on the spatio-temporal graph convolution network model, the node set is projected into a space attention space, the space attention weight of each node in the node set is calculated according to the space attention space projection of the node set, and the space attention matrix is obtained based on the space attention weight of each node;
[0078] Based on the graph convolution method, the space attention matrix is processed to extract the space features of each node in the node set;
[0079] Based on the time attention matrix, the time features of each node in the node set are extracted;
[0080] Obtain a spatio-temporal fusion feature of each node in the node set based on a spatial feature of each node in the node set and a temporal feature of each node.
[0081] The time attention matrix is obtained based on the temporal attention weight of each node, and includes:
[0082] The telemetry data of each node in the node set is linearly processed by using a learnable temporal linear transformation matrix to project the node set to a temporal attention space;
[0083] The time similarity matrix between each node is calculated according to the temporal attention space projection of the node set;
[0084] The time bias item matrix is added to the time similarity matrix between each node, and the time attention weight matrix is generated by using an activation function;
[0085] The time attention weight matrix is linearly processed by using a learnable temporal weight matrix to calculate the temporal attention weight of each node in the node set, and the time attention matrix is obtained based on the temporal attention weight of each node;
[0086] Correspondingly, the node set is projected to a spatial attention space based on the spatio-temporal graph convolution network model, the spatial attention weight of each node in the node set is calculated according to the spatial attention space projection of the node set, and the spatial attention matrix is obtained based on the spatial attention weight of each node, and includes:
[0087] The node set is weighted by using the time attention matrix to obtain a time weighted feature matrix;
[0088] The node set is linearly processed by using a learnable spatial linear transformation matrix to project the node set to a spatial attention space;
[0089] The spatial similarity matrix between each node is calculated according to the spatial attention space projection of the node set;
[0090] The spatial bias item matrix is added to the spatial similarity matrix between each node, and the spatial attention weight matrix is generated by using an activation function;
[0091] The spatial attention weight matrix is linearly processed by using a learnable spatial weight matrix to calculate spatial attention weights of each node in the node set respectively, and a spatial attention matrix is obtained based on the spatial attention weights of each node.
[0092] The spatial attention matrix is processed based on a graph convolution method to extract spatial features of each node in the node set, including:
[0093] The spatial attention matrix is normalized to obtain a normalized spatial attention matrix;
[0094] The normalized spatial attention matrix is processed by graph convolution using Chebyshev polynomials to extract spatial features of each node in the node set;
[0095] Correspondingly, based on the temporal attention matrix, the temporal features of each node in the node set are extracted, including:
[0096] The temporal attention matrix is processed by two-dimensional convolution to extract the temporal features of each node in the node set.
[0097] In one possible implementation, after obtaining the spatio-temporal fusion features of each node in the node set based on the spatial features of each node and the temporal features of each node in the node set in step S300, the method further includes:
[0098] The node set is constructed by a multi-head graph attention mechanism based on a graph neural network model;
[0099] The graph attention weights of each node in the node set are calculated by using the multi-head graph attention mechanism;
[0100] Each node in the node set is processed according to the graph attention weights of each node in the node set to calculate graph attention features by each attention head in the multi-head graph attention mechanism;
[0101] The graph attention features calculated by each attention head are concatenated with the spatio-temporal fusion features of each node in the node set to obtain power grid node features.
[0102] S400. The power grid node features are input into a one-dimensional convolutional neural network, and a one-dimensional convolution kernel is used for node coverage to map the power grid node features to power grid edge features;
[0103] In step S400, the power grid node features are input into a one-dimensional convolutional neural network, and a one-dimensional convolution kernel is used for node coverage to map the power grid node features to power grid edge features, including:
[0104] Topologically arrange the power grid node features based on the graph model to form a feature input matrix;
[0105] Perform a convolution operation on the feature input matrix;
[0106] Slide and cover each adjacent node in the feature input matrix with a convolution kernel to extract power grid edge feature corresponding to the power grid node feature.
[0107] S500. Reduce the dimension of the power grid edge feature, and output the probability value of each edge in the edge set being a fault section, and select the corresponding edge with the maximum probability value as the power grid fault positioning result.
[0108] Through the power grid fault positioning method based on the graph neural network in the embodiment, first, the actual power grid line is acquired, and a graph model is generated; then, the telemetry data and the remote signaling data corresponding to each electrical operating device are acquired respectively and stored in each node correspondingly; then, the attention-based spatio-temporal graph convolution network model and the graph neural network model are trained, and the power grid node features are extracted; subsequently, the power grid node features are input into a one-dimensional convolution neural network for node covering, so as to map the power grid node features to the power grid edge features; finally, the probability value of each edge in the edge set being a fault section is output, and the corresponding edge with the maximum probability value is selected as the power grid fault positioning result. The scheme provided by the application not only has strong robustness to topology changes, high-resistance faults and incomplete data scenes, but also significantly improves the accuracy of power distribution network fault positioning.
[0109] The second aspect of the embodiment provides a hardware system for implementing the power grid fault positioning method based on the graph neural network in the first aspect of the embodiment, and the hardware system comprises:
[0110] An acquisition unit is configured to acquire nuclear magnetic scanning parameters, wherein the nuclear magnetic scanning parameters comprise a scanning repetition number.
[0111] A parameter determination unit is configured to determine, according to the nuclear magnetic scanning parameters, a flip time corresponding to each scanning dimension in a plurality of scanning dimensions, an echo train length corresponding to each flip time, a plurality of gradient rotation angle values, and a total scanning time corresponding to each gradient rotation angle value, wherein the echo train lengths corresponding to the flip times decrease in turn.
[0112] The scanning unit is configured to generate a gradient magnetic field corresponding to each gradient rotation angle value based on the plurality of gradient rotation angle values, and to perform pulse excitation on the object to be measured in each gradient magnetic field according to the scan repetition number, the inversion time corresponding to each scan dimension, and the echo train length corresponding to each inversion time, so as to acquire the nuclear magnetic resonance signal of the object to be measured in each gradient magnetic field after the pulse excitation, wherein the gradient of the gradient magnetic field corresponding to any gradient rotation angle value is constant, and the magnetic field duration is greater than the total scan time corresponding to the any gradient rotation angle value.
[0113] The nuclear magnetic two-dimensional spectrum generation unit is configured to generate a nuclear magnetic two-dimensional spectrum of the object to be measured according to each nuclear magnetic resonance signal.
[0114] The working process, working details and technical effects of the system provided by the embodiment can be referred to the first aspect of the embodiment, and will not be repeated here.
[0115] The third aspect of the embodiment provides an electronic device, which comprises a memory, a processor and a transceiver connected in sequence, wherein the memory is configured to store a computer program, the transceiver is configured to receive and send messages, and the processor is configured to read the computer program and execute the power grid fault positioning method based on the graph neural network as described in the first aspect of the embodiment.
[0116] For example, the memory can include, but is not limited to, a random access memory (RAM), a read-only memory (ROM), a flash memory, a first-in first-out memory (FIFO) and / or a first-in last-out memory (FILO), etc.; specifically, the processor can include one or more processing cores, such as a 4-core processor, an 8-core processor, etc. The processor can be implemented in at least one of a hardware form of a DSP (Digital Signal Processing), a FPGA (Field-Programmable Gate Array) and a PLA (Programmable Logic Array), and the processor can also include a main processor and a coprocessor. The main processor is a processor for processing data in an awake state, also known as a CPU (Central Processing Unit); the coprocessor is a low-power processor for processing data in a standby state.
[0117] In some embodiments, the processor can be integrated with a GPU (Graphics Processing Unit) that is responsible for rendering and drawing the content required to be displayed on the display screen, for example, the processor can not be limited to a microprocessor of STM32F105 series, a RISC (reduced instruction set computer) microprocessor, an X86 architecture processor, or a processor integrated with an embedded NPU (neural-network processing units); the transceiver can be but not limited to a WIFI wireless transceiver, a Bluetooth wireless transceiver, a GPRS (General Packet Radio Service) wireless transceiver, a ZigBee wireless transceiver, a 3G transceiver, a 4G transceiver, and / or a 5G transceiver, etc. In addition, the device can also include but not limited to a power module, a display screen, and other necessary components.
[0118] The working process, working details and technical effects of the electronic device provided by the embodiment can be referred to the first aspect of the embodiment, and will not be repeated here.
[0119] The embodiment provides a computer program product containing instructions, when the instructions are executed on a computer, the computer executes the power grid fault positioning method based on a graph neural network as described in the first aspect of the embodiment, wherein the computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices.
[0120] Finally, it should be noted that: the above only describes the preferred embodiments of the present application and is not used to limit the protection scope of the present application. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.
Claims
1. A power grid fault location method based on a graph neural network, characterized in that, The method comprises the following steps: acquiring an actual power grid line and generating a graph model according to the actual power grid line, wherein the graph model comprises a node set and an edge line set, the node set is used to represent various electrical operating devices, and the edge line set is used to represent power line sections; acquiring telemetry data and telesignaling data corresponding to each electrical operating device respectively, and storing the telemetry data and the telesignaling data corresponding to each electrical operating device into each node of the node set respectively, wherein the telemetry data comprises voltage signals and current time sequence signals of the corresponding electrical operating device, and the telesignaling data comprises discrete action information of the corresponding electrical operating device; training a spatio-temporal graph convolution network model based on attention and a graph neural network model based on the node set, and processing the node set according to the spatio-temporal graph convolution network model based on attention and the graph neural network model to extract power grid node features; inputting the power grid node features into a one-dimensional convolutional neural network, and performing node coverage by using a one-dimensional convolution kernel to map the power grid node features to power grid edge line features; performing dimension reduction on the power grid edge line features, and outputting probability values of each edge line in the edge line set being a fault section, and selecting a corresponding edge line with the largest probability value as a power grid fault positioning result; training a spatio-temporal graph convolution network model based on attention and a graph neural network model based on the node set, comprising: acquiring telemetry data of each node in the node set; forming a first training set with the telemetry data of each node; training a spatio-temporal graph convolution network model based on attention based on the first training set; acquiring telesignaling data of each node in the node set; forming a second training set with the telesignaling data of each node; training a graph neural network model based on the second training set; processing the node set according to the spatio-temporal graph convolution network model based on attention and the graph neural network model to extract power grid node features, comprising: projecting the node set to a time attention space based on the spatio-temporal graph convolution network model, calculating time attention weights of each node in the node set according to the time attention space projection of the node set, and obtaining a time attention matrix based on the time attention weights of each node; projecting the node set to a space attention space based on the spatio-temporal graph convolution network model, calculating space attention weights of each node in the node set according to the space attention space projection of the node set, and obtaining a space attention matrix based on the space attention weights of each node; processing the space attention matrix based on a graph convolution method to extract space features of each node in the node set; extracting time features of each node in the node set based on the time attention matrix; obtaining spatio-temporal fusion features of each node in the node set based on the space features of each node in the node set and the time features of each node in the node set; after obtaining the spatio-temporal fusion features of each node in the node set based on the space features of each node in the node set and the time features of each node in the node set, the method further comprises: Based on the graph neural network model, a multi-head graph attention mechanism is constructed for the node set; The graph attention weight of each node in the node set is calculated by using the multi-head graph attention mechanism; According to the graph attention weight of each node in the node set, the nodes in the node set are processed to calculate the graph attention feature by each attention head in the multi-head graph attention mechanism; The graph attention features calculated by the attention heads are spliced with the spatio-temporal fusion features of each node in the node set to obtain the power grid node features; The node set is projected into a time attention space, and the time attention weight of each node in the node set is calculated according to the time attention space projection of the node set, and a time attention matrix is obtained based on the time attention weight of each node, including: The telemetry data of each node in the node set is linearly processed by using a learnable time linear transformation matrix to project the node set into a time attention space; According to the time attention space projection of the node set, a time similarity matrix between nodes is calculated; A time bias item matrix is added to the time similarity matrix between nodes, and a time attention weight matrix is generated by using an activation function; The time attention weight matrix is linearly processed by using a learnable time weight matrix to calculate the time attention weight of each node in the node set, and a time attention matrix is obtained based on the time attention weight of each node. Correspondingly, based on the spatio-temporal graph convolution network model, the node set is projected into a spatial attention space, and the spatial attention weight of each node in the node set is calculated according to the spatial attention space projection of the node set, and a spatial attention matrix is obtained based on the spatial attention weight of each node, including: The nodes in the node set are weighted processed by using the time attention matrix to obtain a time weighted feature matrix; The nodes in the node set are linearly processed by using a learnable spatial linear transformation matrix to project the node set into a spatial attention space; According to the spatial attention space projection of the node set, a spatial similarity matrix between nodes is calculated; A spatial bias item matrix is added to the spatial similarity matrix between nodes, and a spatial attention weight matrix is generated by using an activation function; The spatial attention weight matrix is linearly processed by using a learnable spatial weight matrix to calculate the spatial attention weight of each node in the node set, and a spatial attention matrix is obtained based on the spatial attention weight of each node.
2. The graph neural network-based power grid fault location method according to claim 1, characterized in that, Based on the graph convolution method, the spatial attention matrix is processed to extract the spatial features of each node in the node set, including: The spatial attention matrix is normalized to obtain a normalized spatial attention matrix; The normalized spatial attention matrix is graph-convoluted by using Chebyshev polynomials to extract the spatial features of each node in the node set. Correspondingly, based on the time attention matrix, the time characteristics of each node in the node set are extracted, and the time characteristics of each node in the node set are extracted. The time attention matrix is subjected to two-dimensional convolution processing to extract the time characteristics of each node in the node set.
3. The graph neural network-based power grid fault location method according to claim 1, characterized in that, The power grid node characteristics are input into a one-dimensional convolutional neural network, and a one-dimensional convolution kernel is used for node coverage to map the power grid node characteristics to power grid edge line characteristics, including: Based on the graph model, the power grid node characteristics are topologically arranged to form a feature input matrix; The feature input matrix is subjected to convolution operation; Each adjacent node in the feature input matrix is covered by a convolution kernel to extract the power grid edge line characteristics corresponding to the power grid node characteristics.
4. A power grid fault location system based on graph neural networks, characterized in that, The method is applied to realize the power grid fault location method based on the graph neural network as claimed in any one of claims 1-3, including: A data acquisition unit is configured to acquire an actual power grid line and generate a graph model based on the actual power grid line, wherein the graph model includes a node set and an edge line set, the node set is used to represent each electrical operating device, and the edge line set is used to represent a power line section; A data storage unit is configured to acquire telemetry data and telesignaling data corresponding to each electrical operating device, respectively, and store the telemetry data and telesignaling data corresponding to each electrical operating device into each node of the node set, respectively, wherein the telemetry data includes voltage signals and current time sequence signals of the corresponding electrical operating device, and the telesignaling data includes discrete action information of the corresponding electrical operating device; A feature extraction unit is configured to train an attention-based spatio-temporal graph convolution network model and a graph neural network model based on the node set, and process the node set based on the attention-based spatio-temporal graph convolution network model and the graph neural network model to extract power grid node characteristics; and input the power grid node characteristics into a one-dimensional convolutional neural network and use a one-dimensional convolution kernel for node coverage to map the power grid node characteristics to power grid edge line characteristics; A fault location unit is configured to reduce the dimension of the power grid edge line characteristics and output probability values of each edge line in the edge line set being a fault section, and select the corresponding edge line with the largest probability value as the power grid fault location result.
5. An electronic device, comprising: including: The memory, the processor and the transceiver are sequentially connected in communication, wherein the memory is configured to store a computer program, the transceiver is configured to receive and transmit messages, and the processor is configured to read the computer program and execute the power grid fault location method based on the graph neural network as claimed in any one of claims 1-3.
6. A computer program product comprising instructions, characterized in that, When the instructions are run on the computer, the computer executes the power grid fault location method based on the graph neural network as claimed in any one of claims 1-3.