Power grid topology fault location system based on graph neural network

Abstract

The present invention relates to the field of neural network technology, and further to a power grid topology fault location system based on a graph neural network. The system comprises: a power grid topology feature extraction unit, which is used to regard the power grid as a graph network and perform real-time data collection on each node of the power grid, wherein each element in the feature matrix represents the electrical characteristics of a certain node of the power grid at a certain time; a multi-layer graph neural network model unit, which is used to use the feature matrix as the input of the first layer as the feature extraction result; a fault identification and location unit, which is used to calculate the fault probability distribution based on the feature extraction result; and the credibility of the calculated fault location. The present invention improves the system's modeling capability for complex power grid topologies, enhances the real-time and robustness of fault detection, and realizes the precise location of the fault area and location.

Description

Technical Field

[0001] The present invention belongs to the technical field of neural networks, and in particular relates to a power grid topology fault location system based on a graph neural network. Background Art

[0002] Traditional methods for detecting and locating power grid faults mainly rely on signal analysis, model prediction, and data-driven machine learning algorithms. Common fault detection technologies include protection devices based on voltage and current waveforms, state estimation methods based on phasor measurements, and methods based on signal processing and transform domain analysis. Although these methods have achieved certain results in practical applications, with the continuous expansion of the scale of power grids and the increase in operational complexity, traditional detection methods face challenges in the following aspects:

[0003] Fault propagation paths under complex topological structures are difficult to model accurately: In traditional power grids, the connection relationship between nodes is usually simple, so the fault propagation path can be described by some fixed transmission equations. However, the topological structure of modern power grids is increasingly complex, with a large number of branches and redundant connections. Traditional mathematical models are difficult to effectively capture the complex node-to-node relationships in the power grid. In particular, when multiple point faults or complex conduction faults occur in the power grid, the propagation model based on linear equations in traditional methods often fails and is difficult to accurately reflect the fault propagation path and impact range.

[0004] Existing methods lack real-time and robustness: Real-time detection of power grid faults is very important, because rapid fault location and repair can significantly reduce power supply interruption time and reduce economic losses. Many existing methods have bottlenecks in real-time fault detection, mainly because these methods require a lot of computing resources to process large-scale data of the power grid, or rely on preset rules and models in fault mode analysis. In addition, when faced with uncertain load changes or external interference, traditional model prediction methods are very sensitive to the noise of input data and it is difficult to maintain high detection accuracy in a noisy environment.

[0005] Insufficient ability to capture dynamic electrical features: The electrical parameters of the power grid, including voltage, current, power, frequency, etc., usually change dynamically with changes in load or changes in the state of the grid. This change is sometimes a slow trend change, and sometimes a drastic short-term fluctuation. Many existing fault detection methods mainly focus on the analysis of static features, or rely on linear feature extraction methods, which makes it difficult to effectively capture these dynamically changing features in the power grid, especially when a power grid fault occurs, the electrical parameters will show highly nonlinear and drastic changes. Traditional linear feature analysis methods are often unable to accurately identify the cause and location of the fault in the case of such nonlinear feature changes. Summary of the invention

[0006] The main purpose of the present invention is to provide a power grid topology fault location system based on graph neural network. The present invention improves the system's modeling capability of complex power grid topologies, enhances the real-time and robustness of fault detection, and realizes the precise location of fault areas and positions.

[0007] In order to solve the above problems, the technical solution of the present invention is achieved as follows:

[0008] A power grid topology fault location system based on graph neural network, the system includes: a power grid topology feature extraction unit, which is used to regard the power grid as a graph network, collect real-time data for each node of the power grid, record the real-time electrical measurement value of each node, and construct a feature matrix, each element in the feature matrix represents the electrical characteristics of a node of the power grid at a certain time; a multi-layer graph neural network model unit, which is used to use the feature matrix as the input of the first layer, combine the adjacency matrix and degree matrix of the power grid, and after multi-layer feature extraction of the preset multi-layer graph neural network model, output the result of the last layer as the feature extraction result; a fault identification and location unit, which is used to construct a fault propagation dynamics model based on the feature extraction result to obtain a fault feature field; use the fault feature field to calculate the fault probability distribution; determine the fault area and fault boundary according to the fault probability distribution; calculate the fault propagation distribution according to the fault area and the fault feature field; calculate the fault location and the credibility of the calculated fault location according to the fault propagation distribution.

[0009] Furthermore, the electrical measurements include: voltage amplitude, current amplitude, voltage phase angle, current phase angle, active power, reactive power, frequency deviation and node impedance.

[0010] Furthermore, the feature matrix is:

[0011]

[0012] Where X(t) is the characteristic matrix of the power grid at time t; N is the total number of nodes in the power grid; V k (t) is the voltage amplitude of node k at time t; cos(θ k (t)) is the cosine value of the voltage phase angle of node k at time t; I k (t) is the current amplitude of node k at time t; sin(φ k (t)) is the sine value of the current phase angle of node k at time t; P k (t) is the active power of node k at time t; Q k (t) is the reactive power of node k at time t; S base is the reference value of power, used to normalize active power and reactive power; Δf k(t) is the frequency deviation of node k at time t, that is, the actual frequency relative to the reference frequency f 0 The deviation of Z k (t) is the impedance of node k at time t; Z base is the reference value of impedance; Δt is the time step, which indicates the time difference between two times; λ is the attenuation parameter, which controls the rate at which the characteristic matrix decays over time and is the set value.

[0013] Furthermore, the preset multi-layer graph neural network model has a total of l layers, and the feature extraction result is obtained by the following formula:

[0014]

[0015] Among them, H l is the result of the last layer, as the result of feature extraction; D is the degree matrix of the power grid; A is the adjacency matrix of the power grid; W l is the weight of the lth layer; σ(·) is the activation function; ⊙ is the Hadamard product; H l-1 is the result of the l-1th layer; for the input layer, H 0 =X(t).

[0016] Furthermore, based on the feature extraction results, the formula for constructing the fault propagation dynamics model is:

[0017]

[0018] Where k is the subscript integer index; H k Represents the result of the kth layer of the multi-layer graph neural network model; represents the gradient operator; is the Laplace operator; F(t) is the fault characteristic field at time t.

[0019] Furthermore, the fault characteristic field is used to calculate the fault probability distribution P(r, t) through the following formula:

[0020]

[0021] Where P(r, t) represents the fault probability distribution at location r and time t; Ω is the power grid domain.

[0022] Furthermore, the fault area and fault boundary are determined according to the fault probability distribution through the following formula:

[0023]

[0024] Among them, R fault is the fault area; B fault is the fault boundary; η th is the probability threshold; ξth is the probability gradient threshold; n is the boundary normal vector; s is the boundary integral variable, which represents the position of each tiny line segment on the boundary of the fault area.

[0025] Furthermore, according to the fault area and fault characteristic field, the fault propagation distribution is calculated by the following formula:

[0026]

[0027] Among them, T fault (t) is the fault propagation distribution at time t; τ is the time integral variable; is the tensor product.

[0028] Furthermore, the fault location and the credibility of the calculated fault location are calculated by the following formula:

[0029]

[0030] Among them, L final is the fault location; C reliability is the fault location L final The credibility of; T is the total time.

[0031] The power grid topology fault location system based on graph neural network of the present invention has the following beneficial effects: the present invention models the topological structure of the power grid through graph neural network, and can fully capture the complex relationship between nodes in the process of feature extraction. The graph neural network uses the adjacency matrix and degree matrix of the power grid as input, fuses the node features layer by layer, and realizes the modeling of the global topological structure through feature propagation and nonlinear activation. This method can not only capture the global topological relationship of the power grid, but also improve the expression ability of local complex relationships through multi-layer feature extraction, so as to more accurately reflect the fault propagation path of the power grid. Compared with the traditional method, the present invention can show higher modeling ability when facing large-scale and complex topological power grids, and effectively solves the problem that the fault propagation path is difficult to describe in the case of multiple nodes and multiple connections. The electrical parameters of the power grid are usually affected by external interference and noise. The traditional fault detection method is often sensitive to the noise of the input data, which is easy to cause false alarms or missed alarms. The present invention introduces the feature extraction mechanism of the graph neural network, deeply fuses the electrical characteristics and adjacency relationships of the nodes in the process of feature propagation, and effectively improves the immunity of the system to noise. In addition, through the dynamic time decay processing of the feature matrix, the system can adaptively adjust the contribution of features at different time points to the current moment, thereby enhancing the sensitivity to dynamic changes and noise resistance. Different from the traditional static analysis method, the present invention uses a time-related feature fusion strategy to better capture the dynamic change characteristics of nodes in the power grid, and improve the detection accuracy and robustness in a complex power grid environment. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 A schematic diagram of the system structure of a power grid topology fault location system based on a graph neural network provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0033] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0034] Example 1, reference Figure 1 : A power grid topology fault location system based on graph neural network, the system includes: a power grid topology feature extraction unit, which is used to regard the power grid as a graph network, collect real-time data for each node of the power grid, record the real-time electrical measurement value of each node, and construct a feature matrix, each element in the feature matrix represents the electrical characteristics of a node of the power grid at a certain time; a multi-layer graph neural network model unit, which is used to take the feature matrix as the input of the first layer, combine the adjacency matrix and degree matrix of the power grid, and after multi-layer feature extraction of the preset multi-layer graph neural network model, output the result of the last layer as the feature extraction result; a fault identification and location unit, which is used to construct a fault propagation dynamics model based on the feature extraction result to obtain a fault feature field; use the fault feature field to calculate the fault probability distribution; determine the fault area and fault boundary according to the fault probability distribution; calculate the fault propagation distribution according to the fault area and the fault feature field; calculate the fault location and the credibility of the calculated fault location according to the fault propagation distribution.

[0035] Specifically, the power grid topology feature extraction unit regards the entire power grid as a graph network, where the graph consists of each node in the power grid and the connection relationship between the nodes. Each node in the system corresponds to an electrical device with physical properties, such as a substation, a distribution room or a switch station, and the connection relationship between the nodes is composed of physical connections such as power lines or cables. In order to better describe this graph network, the power grid topology feature extraction unit collects real-time data from each node, records the electrical measurement value of each node, including key information such as voltage, current, phase angle, power, etc., and constructs these data into a feature matrix. Each element in the feature matrix represents the specific electrical characteristics of a node in the power grid at a certain point in time. Through this matrix form, the state of each node in the power grid can be digitized and structured, which is convenient for the subsequent processing of the graph neural network. Next, the unit uses the feature matrix as input and uses the adjacency matrix and degree matrix to characterize the connection relationship between nodes. The adjacency matrix is ​​used to represent the direct connection between nodes in the power grid, that is, which nodes have power connections, and the degree matrix is ​​used to characterize the number of connections of each node, thereby measuring the importance of the node and the strength of the association between nodes. After constructing the feature matrix, adjacency matrix and degree matrix, the power grid topology feature extraction unit inputs these matrices together into the preset multi-layer graph neural network model. The core idea of ​​the graph neural network is to update node features through layer-by-layer feature propagation and aggregation. Specifically, each layer of the graph neural network gradually updates the features of a node by combining and weighted averaging the features of a node with the features of its neighboring nodes. In this process, the way the features are propagated depends not only on the features of the node itself, but also on the connection relationship between the nodes. In other words, the electrical characteristics of each node in the power grid will be affected by the information of the neighboring nodes during the process of feature propagation. This layer-by-layer update of features can fully capture the global characteristics of the power grid topology.

[0036] In the process of feature extraction, in order to improve the accuracy of feature expression, the power grid topology feature extraction unit not only adopts a simple weighted average strategy, but also combines nonlinear activation functions and normalization operations to better adapt to the complex and changeable state of the power grid. Specifically, by introducing nonlinear activation functions, the changes in node features are no longer just the result of linear superposition, but can reflect more complex relationships between nodes. The normalization operation can prevent the distortion or overflow of features during multi-layer transmission, thereby ensuring the stability and effectiveness of the features. After multi-layer feature extraction, the output of the final layer will be used as a comprehensive feature expression of the power grid. These features contain the feature information of each node after considering the influence of its neighboring nodes, which can more comprehensively describe the global state of the power grid. In the principle of the entire power grid topology feature extraction unit, the core of feature extraction lies in the repeated propagation and fusion of power grid node features and connection relationships through multi-layer graph neural networks, thereby realizing the dynamic update of node features and accurate modeling of global features. Compared with the traditional power grid topology modeling method, this unit can not only capture the local electrical characteristics of a single node, but also reflect the topological structure characteristics between nodes through the adjacency matrix and degree matrix, realizing a global description of the entire power grid. Therefore, the design of the power grid topology feature extraction unit innovatively introduces the graph neural network model, incorporating the topological relationship of the power grid into the feature extraction process, which can better reflect the state of the power grid and its potential fault characteristics at the feature level.

[0037] In the multi-layer graph neural network model unit, the most critical part is the multi-layer feature propagation mechanism. Specifically, the graph neural network gradually updates the features of each node through multiple stacked neural network layers. Each layer of the neural network generates a new feature representation by weighted fusion of the feature information of the current layer with the feature information of the neighboring nodes. This feature fusion process not only depends on the feature value of the single node itself, but also propagates the features of the neighboring nodes based on the adjacency relationship between the nodes, that is, the direct connectivity between the nodes is characterized in the form of an adjacency matrix. Therefore, in each layer of the multi-layer graph neural network, the system aggregates the current feature vector of each node with the features of all its connected nodes, and normalizes it in combination with the degree matrix. The purpose of this process is to ensure that the influence between nodes can be appropriately adjusted during the feature update process, thereby avoiding feature distortion caused by degree differences.

[0038] Through multi-layer feature extraction, each layer of the neural network updates and reconstructs the features of the node from a specific perspective, and as the number of network layers increases, the feature vector of the node will gradually integrate information from more distant neighbors. This layer-by-layer propagation mechanism makes the final feature representation of each node not only contain its own electrical characteristics, but also integrate the state information of all other nodes connected to it, thus forming a global description of the entire power grid topology. In particular, the complex node relationships in the power grid topology, such as the redundancy of power transmission paths, loop connections, etc., can be fully modeled and captured through the feature propagation mechanism of the multi-layer graph neural network. This enables the system to maintain high accuracy and stability when dealing with large-scale nodes in the power grid. Another important part of the multi-layer graph neural network model unit is the introduction of nonlinear activation functions. After each layer of feature propagation, the nonlinear activation operation of the node features can improve the model's ability to express complex feature patterns. Traditional linear activation methods are often only applicable to linear combinations of simple features, while in complex topological structures such as power grids, the connection relationships and mutual influences between different nodes may have nonlinear characteristics. Therefore, the introduction of nonlinear activation functions helps the system better capture these complex node feature changes. In addition, the design of normalization operation also plays a crucial role in multi-layer graph neural network. During the feature propagation process, there may be significant differences in the connection strength between nodes, and these differences may lead to distortion of feature updates if they are not regulated. Through the normalization operation, the system can dynamically adjust the feature propagation process of each node, making the feature extraction more stable and accurate. Through the above multi-layer feature propagation and nonlinear activation process, the multi-layer graph neural network model unit gradually builds a feature representation with global topology awareness. The output of the final layer represents the deep-level features of each node in the power grid. It integrates the features of all adjacent nodes and the global information of the power grid topology, laying the foundation for subsequent fault identification and location. Compared with traditional fault detection methods, this unit realizes global modeling of the power grid state through multi-layer feature extraction, which can more accurately characterize the state changes of each node in the power grid before and after the fault occurs. At the same time, the feature propagation mechanism based on the adjacency matrix and degree matrix enables the system to flexibly adapt to the topological changes of the power grid, thereby improving the robustness of fault location. In addition, it is worth emphasizing that in the actual fault location process, the design of the multi-layer graph neural network model unit also considers the interpretability of the feature extraction results. Since the feature updates of nodes in each layer of the network are based on the features of their neighboring nodes, the final node feature representation can reflect the dependencies and information propagation paths between nodes in the power grid. In this way, when a power grid fails, the system can not only identify the fault by extracting deep features, but also track the propagation path of the fault and identify potential fault areas and key nodes by analyzing the feature propagation process.This combination of global topology modeling and fault propagation path enables the multi-layer graph neural network model unit to demonstrate strong explanatory power and flexible adaptability in power grid fault location.

[0039] The working principle of the fault identification and location unit can be elaborated in detail from the following aspects. First, the unit uses the grid feature representation output by the previous unit, that is, the output features of the last layer of the graph neural network, as input data. These features have integrated the state information of each node and its adjacent nodes, representing the global characteristics of the grid. Based on these features, the fault identification and location unit will construct a dynamic model of fault propagation to simulate the propagation of faults between nodes in the grid. Traditional grid fault analysis methods usually only consider the independent state changes of nodes, while ignoring the dynamic mutual influence in the grid topology. This unit overcomes this defect by introducing a fault propagation dynamic model, treating the fault as a dynamic propagation process, and simulating the fault propagation path based on the grid topology and node characteristics. In the fault propagation dynamic model, the unit combines the adjacency matrix of the grid and the degree information of the node to simulate the fault propagation process. Specifically, when a node is abnormal, the state change of the node may affect the adjacent nodes through the connection relationship, thereby causing the fault to spread further. In order to accurately capture this propagation behavior, the fault propagation dynamics model describes the interaction between nodes through a set of parameterized propagation rules, which are dynamically adjusted according to the characteristic information of each node and the topological relationship of the power grid. Therefore, the unit can quickly identify the direction and range of fault propagation when a fault occurs, and infer the area that may be affected based on the propagation path.

[0040] On the basis of the fault propagation model, the fault identification and location unit further introduces the concept of fault characteristic field. The fault characteristic field is a spatial distribution field that reflects the fault state. It represents the distribution of faults in the power grid by spatially mapping the characteristics of each node in the power grid. The construction process of the fault characteristic field is to combine the output results of the fault propagation model with the node characteristics to generate a global characteristic field that reflects the fault state. This characteristic field can not only show the location of the fault, but also describe the propagation trend of the fault in the power grid. By analyzing the fault characteristic field, the system can further calculate the probability of the fault. The probability distribution of the fault indicates the possibility of each node failing. The higher the probability of the node, the greater the possibility of the node failing. In this way, the fault identification and location unit can identify the possible origin and main impact area of ​​the fault at the early stage of the fault. After obtaining the fault probability distribution, the fault identification and location unit will further determine the fault area and fault boundary based on the probability information. The determination of the fault area is based on the analysis of the fault probability distribution. The nodes with probability values ​​higher than a certain threshold are selected as nodes in the fault area, thereby identifying the main impact range of the fault. At the same time, the system will calibrate the fault boundary according to the fault propagation path simulated by the fault propagation model, that is, the nodes that may be affected by the fault but have not yet completely failed. In this way, the unit can accurately define the impact area of ​​the fault and provide a strong basis for further handling of the fault.

[0041] After the fault area and fault boundary are determined, the fault identification and location unit will further calculate the fault propagation distribution according to the fault characteristic field and the results of fault propagation. The fault propagation distribution represents the specific path and strength of the fault spreading from the starting node to other nodes. To achieve this goal, the unit will combine the adjacency relationship of the power grid and the characteristic information of each node, and use the preset propagation rules to simulate the fault diffusion process. In this process, the system will dynamically adjust the node weights and connection relationships on the propagation path to accurately simulate the propagation behavior of the fault. This propagation distribution calculation can not only help the system better understand the propagation mode of the fault, but also provide a reference for the specific location of the fault. Based on the fault propagation distribution, the final task of the fault identification and location unit is to calculate the location of the fault and the credibility of the fault location. The calculation of the fault location is determined by analyzing the fault characteristic field and the propagation path. Specifically, the system will determine the starting location and main impact range of the fault based on the node status information on the propagation path. In this process, the system will comprehensively consider the fault probability of each node, the strength of the fault propagation path, and the interaction between nodes to accurately determine the specific location of the fault. At the same time, in order to improve the reliability of fault location, the system will also calculate the credibility of the fault location. Credibility indicates the reliability of the fault location. The higher the credibility, the more accurate the fault location. Credibility is calculated by comprehensively analyzing the propagation path, node status, and adjacency to evaluate the confidence of fault location.

[0042] Embodiment 2: Electrical measurements include: voltage amplitude, current amplitude, voltage phase angle, current phase angle, active power, reactive power, frequency deviation and node impedance.

[0043] Specifically, the voltage amplitude and current amplitude represent the voltage and current intensity of the node at a certain moment, respectively. These two basic electrical quantities are the basic elements of fault detection and analysis, and can intuitively reflect the power transmission and consumption of the node. Any significant fluctuation in the voltage or current of a node may be caused by a fault or a change in the electrical characteristics of the node. Therefore, by recording the voltage amplitude and current amplitude of the node, the most basic power characteristic data can be provided to the system. The voltage phase angle and the current phase angle represent the phase difference between the voltage and current relative to the reference signal. The changes in the voltage phase angle and the current phase angle not only reflect the power transmission state of the node, but also reveal the influence of the power factor and reactive power. In the power grid, since the occurrence of a fault may cause a drastic change in the voltage or current phase angle, these phase angle information can be used to further determine the relative relationship and synchronization between nodes. By analyzing the offset of the voltage phase angle and the current phase angle, the system can better infer the cause and propagation path of the fault. The measurement of active power and reactive power is a quantitative description of the node power transmission efficiency and the power flow between nodes. Active power represents the ability of electric energy to actually do work in the power grid, and is the effective power transmitted by the power grid; while reactive power reflects the energy exchange caused by the reactance elements in the power grid, and is usually used for voltage regulation and reactive power compensation of power equipment. Since faults will affect the efficiency of power transmission, by monitoring the changes in the active power and reactive power of the node, the imbalanced state of power distribution in the power grid can be intuitively reflected, thus providing an important reference for fault identification. Frequency deviation refers to the offset of the grid frequency of the node relative to the standard value. The frequency stability of the power grid is directly related to the normal operation of the power grid. When a fault occurs, the frequency of the power grid often fluctuates or deviates. By measuring the frequency deviation of the node, the system can monitor the changes in the grid frequency, so as to promptly discover possible fault sources and abnormal conditions. Node impedance, as a characteristic parameter of the node in the power grid, represents the impedance value of the node during the power transmission process. The size of the node impedance directly affects the transmission effect of the current and the stability of the power grid. When a fault occurs in the power grid, the node impedance often changes significantly. Therefore, by measuring the node impedance, the location of the fault can be more accurately located and the nature of the fault can be analyzed.

[0044] Example 3: The feature matrix is:

[0045]

[0046] Where X(t) is the characteristic matrix of the power grid at time t; N is the total number of nodes in the power grid; V k (t) is the voltage amplitude of node k at time t; cos(θ k (t)) is the cosine value of the voltage phase angle of node k at time t; I k (t) is the current amplitude of node k at time t; sin(φk (t)) is the sine value of the current phase angle of node k at time t; P k (t) is the active power of node k at time t; Q k (t) is the reactive power of node k at time t; S base is the reference value of power, used to normalize active power and reactive power; Δf k (t) is the frequency deviation of node k at time t, that is, the actual frequency relative to the reference frequency f 0 The deviation of Z k (t) is the impedance of node k at time t; Z base is the reference value of impedance; Δt is the time step, which indicates the time difference between two times; λ is the attenuation parameter, which controls the rate at which the characteristic matrix decays over time and is the set value.

[0047] Specifically, the construction of the characteristic matrix first considers the amplitude and phase angle information of voltage and current. By introducing V k (t)·cos(θ k (t)) and I k (t)·sin(φ k (t)) two items, the matrix can effectively combine the voltage amplitude and phase angle cosine value, as well as the current amplitude and phase angle sine value of the node. This expression not only directly reflects the instantaneous voltage and current characteristics of the node, but also reveals the phase difference between nodes through the phase angle information, which is important for identifying the power transmission status of the power grid, analyzing the power factor, and detecting synchronization problems between nodes. Especially when a power grid fault occurs, these voltage and current phase characteristics often mutate, resulting in a phase angle offset. Therefore, in the feature extraction process of the graph neural network, these features can provide an important basis for the propagation path of the fault. The feature matrix also includes the active power and reactive power information of the node. These two items are obtained through P k (t) / S base and Q k (t) / S base After normalization, it is introduced into the matrix. The normalization of power helps to eliminate the dimensional differences caused by differences in power load between different nodes, making the features more unified and stable during the training and propagation of the graph neural network. By combining active power and reactive power, the feature matrix can not only reflect the actual work done by the node, but also reveal the impact of reactive power on the voltage stability of the node, which is crucial for the detection and analysis of power grid faults. Since changes in reactive power are often a precursor to potential power grid failures, the design of this matrix enables the system to capture this power distribution anomaly in the early stages of a fault, and then further analyze the root cause of the fault through feature extraction of a multi-layer graph neural network.

[0048] The introduction of frequency deviation and impedance further enhances the ability of the characteristic matrix to describe the power grid state. k (t) relative to the reference frequency f 0 The normalization process is performed to capture the frequency changes in the power grid when a fault occurs. The frequency stability of the power grid is directly related to its normal operation, and in actual faults, the frequency deviation of the node often becomes a key feature of fault propagation. Therefore, by normalizing the frequency deviation, the feature matrix enables the graph neural network to more easily identify subtle changes in frequency during training and inference. In addition, the node impedance Z k (t) After Z base After normalization, it is introduced into the matrix, which can describe the impedance characteristics of the node during power transmission. The change in impedance not only reflects the stability problem of the node in current conduction, but also reveals the change in the characteristics of the power line after the fault occurs. Therefore, the impedance characteristics are of great reference significance for the judgment of the fault location. The time decay factor exp(-λΔt) of the feature matrix is ​​an innovation in this design. By introducing time-related exponential decay in the feature matrix, the system can dynamically consider the feature changes in different time steps during the training process of the graph neural network, thereby reflecting the influence of the gradual decay of the features over time. This time decay processing can better adapt to the dynamic changes of the power grid in different time periods, while suppressing the influence of irrelevant historical features on the current fault analysis. Through this design, the feature matrix can more accurately reflect the true state of the power grid node at a certain moment, providing a reliable data basis for subsequent fault identification and location. Overall, the feature matrix integrates the amplitude and phase angle of voltage and current, active power and reactive power, frequency deviation and node impedance, and combines time decay processing, so that the graph neural network model of the present invention can effectively extract key information in the dynamic propagation of node features. Compared with traditional methods, the design of this feature matrix not only takes into account the independent characteristics of each node, but also reflects the interaction between nodes through the relationship between electrical parameters. Especially under fault conditions, the correlation between node features often manifests itself as significant changes in the feature matrix. These changes can form a global modeling of power grid faults through feature fusion and propagation of graph neural networks.

[0049] Example 4: The preset multi-layer graph neural network model has a total of l layers, and the feature extraction result is obtained by the following formula:

[0050]

[0051] Among them, H l is the result of the last layer, as the result of feature extraction; D is the degree matrix of the power grid; A is the adjacency matrix of the power grid; W lis the weight of the lth layer; σ(·) is the activation function; ⊙ is the Hadamard product; H l-1 is the result of the l-1th layer; for the input layer, H 0 =X(t).

[0052] Specifically, input feature matrix H 0 =X(t) represents the node characteristics of the power grid at a specific time t. These characteristics include electrical measurements such as voltage amplitude, current amplitude, phase angle, active power, reactive power, frequency deviation and impedance. The introduction of these characteristics ensures that the initial state of each node can fully reflect its electrical characteristics and provide a sufficient information basis for the subsequent layers of the network. Through the multi-layer graph neural network model, the system updates these node features layer by layer, so that the node characteristics of each layer not only depend on the characteristics of the node itself, but also combine the characteristics of its adjacent nodes. It is this multi-level propagation and fusion of features that enables the system to gradually construct a global topological structure expression of the power grid based on the local characteristics of the power grid. In the feature update process of each layer of the network, the adjacency matrix A and the degree matrix D play a key role. The adjacency matrix A represents the connection between the nodes in the power grid. Through this matrix, the network can identify the relationship between each node and its neighboring nodes. The degree matrix D reflects the number of connections of each node. It is used to normalize the node features to prevent the feature information from being too concentrated or dispersed due to too many connections to some nodes during the feature propagation process. Specifically, in the feature update formula The operation is to normalize the adjacency matrix. This symmetric normalization method ensures that during the feature propagation process, the features of each node not only depend on the features of its neighboring nodes, but also on the strength of node connections, while avoiding excessive deviations in feature updates caused by nodes with larger degrees. Through this operation, the system can balance the feature contributions of different nodes and ensure the stability and uniformity of feature propagation.

[0053] It is worth noting that the feature extraction process not only relies on the topological structure between nodes, but also combines the temporal dynamic changes of the electrical characteristics of the nodes. It reflects the rate of change of the feature matrix X(t) over time, that is, the trend of the node's features over time. Through this derivative term, the system can capture the dynamic feature changes of the power grid nodes in a short period of time, thereby providing more real-time information support for fault identification. In the power grid fault location system, the occurrence of a fault is usually accompanied by a sharp change in the electrical characteristics of the node in a short period of time, such as a sudden drop in the voltage amplitude and a sudden change in the current phase angle. Therefore, by introducing the time-varying derivative of the feature, the system can more sensitively capture the feature fluctuations before and after the fault, thereby improving the accuracy of fault identification. In addition, in the feature update formula, the Hadamard product (i.e., element multiplication) is used to combine the time-varying information of the node feature with the feature propagation result. Specifically, the ⊙ operation enables the time derivative term to work together with the topological feature extraction result to form a feature representation that includes both the topological relationship between nodes and the time dynamic change. This combination ensures that each layer of features can not only reflect the mutual influence between nodes, but also dynamically capture the changes of node features over time. Therefore, when extracting features, the network not only considers the spatial structure of the power grid, but also introduces the characteristics of the time dimension, so that the feature extraction results can better reflect the dynamic evolution process of the power grid. This is crucial for the location of power grid faults, because the occurrence of faults is often accompanied by significant changes in spatial and temporal features, and the multi-layer graph neural network model can extract the global fault characteristics of the power grid layer by layer through the dynamic combination of such features. The introduction of the activation function σ(·) is to enhance the nonlinear expression ability of the network. In the process of power grid fault location, the changes in power grid characteristics are usually highly nonlinear, especially when there are complex topological structures and multiple fault sources in the power grid, and linear models are difficult to capture these complex changes. Through nonlinear activation functions, the network can perform nonlinear transformations on features and enhance the ability to distinguish features, thereby better adapting to the needs of identifying and locating complex power grid faults. Finally, after the feature extraction of the l-layer graph neural network, the system obtains the output H of the last layer. l , as the final feature representation of the power grid node. This feature matrix not only combines the topological structure of the power grid and the electrical characteristics of the nodes, but also incorporates the dynamic change information of the time dimension. Through this multi-layer feature propagation and time feature fusion design, the system can accurately locate the node or area where the fault occurs when facing complex power grid faults, and at the same time has strong fault detection robustness.

[0054] Among them, in the present invention, the adjacency matrix A is the adjacency matrix of the entire power grid model; it can adapt to special cases such as double busbar wiring and 3 / 2 wiring, and has universal applicability. In double busbar wiring, there is a circuit breaker between the two busbars; and in 3 / 2 wiring, there are multiple strings of 3 circuit breakers connected between the two busbars, and lines are drawn between the circuit breakers, and there are cases where the lines are directly connected without passing through the busbar. These two types of wiring are different from the commonly used busbar connection circuit breaker and then the line connection, so they cannot be represented by a general adjacency matrix. It is necessary to establish an adjacency matrix that includes special cases such as double busbar wiring and 3 / 2 wiring. The establishment and calculation method of the power grid characteristic matrix containing special cases such as double busbar connection and 3 / 2 connection includes: the first step is to deeply traverse the power grid connection diagram. If a busbar-circuit breaker-circuit breaker branch is found, the node is considered to be 3 / 2 connection, and the node is marked as T1, until the traversal is completed, Q 3 / 2 connection nodes are generated, recorded as T1...TQ; the second step is to deeply traverse the power grid connection diagram. If a busbar-circuit breaker-busbar branch is found, the node is considered to be double busbar connection, and the node is marked as D1, Until the traversal is completed, M dual-bus nodes are generated, denoted as D1...DM; the third step, because T1...TN external connection nodes include buses and lines, while D1...DM only includes buses, T1...TN can be regarded as a subnet, that is, the submatrix A1 of the adjacency matrix A; D1...DM is regarded as the bus subnet A2; the fourth step, when calculating the connectivity through the adjacency matrix and performing Boolean multiplication, the 3 / 2 connection is calculated as a single node Ti; the dual-bus connection is calculated as a single node Dj. When the calculation is completed, the single node is converted into its own submatrix A2(i) and A2(j). The fifth step, use Boolean addition to add the converted submatrices. If each position of the added matrix is ​​not 0, it is connected.

[0055] Example 5: Based on the feature extraction results, the formula for constructing the fault propagation dynamics model is:

[0056]

[0057] Where k is the subscript integer index; H k Represents the result of the kth layer of the multi-layer graph neural network model; represents the gradient operator; is the Laplace operator; F(t) is the fault characteristic field at time t.

[0058] Specifically, in this formula, the term on the left side represents the rate of change of the fault feature field F(t) over time. This term reflects the dynamic evolution of the fault feature in the power grid and represents the propagation trend of the fault at time t. The two parts of the formula on the right describe the dynamic behavior of the fault feature field from the perspective of spatial topology and node feature changes. First, the first part of the formula Represents the characteristic result H l The Laplace operator is usually used to describe the diffusion phenomenon of the field. In this case, it reflects the spatial diffusion of fault characteristics between nodes. In other words, by applying the Laplace operator, the formula can capture the mutual influence between node characteristics, especially when a node fails, the characteristic changes of the node may be propagated to other nodes through the topological connection relationship of the power grid. The Laplace operator can accurately simulate the diffusion of this characteristic, thereby reflecting the propagation process of the fault in the power grid. Since the topological structure of the power grid is often complex and the fault propagation path may also be nonlinear, this mathematical treatment can better simulate the diffusion behavior of the fault in space. Part II It represents the interaction and gradient change of each layer feature in the multi-layer graph neural network. The gradient operator here is Indicates the changing trend of the k-th layer feature between nodes, and the multiplication operation The purpose of this design is to describe the cumulative effect of node feature changes and how these changes are associated with the direction of fault propagation through the output features of the multi-layer graph neural network. k contains the topological information between the node and its neighbors, and the gradient operator The gradient of node features changing over time or space is revealed. Therefore, by combining these hierarchical features together, the changes in node features at different levels can be captured, thereby building a comprehensive fault propagation model.

[0059] Specifically, when a fault occurs in the power grid, the electrical characteristics of some nodes will first become abnormal, and these abnormal changes will gradually spread to other nodes connected to the faulty node through the topological structure of the power grid. This propagation phenomenon depends not only on the characteristic changes of the node itself, but also on the topological relationship of the power grid. The term is used to describe the cumulative effect of fault propagation. kBy summing and combining its gradient changes, the system can capture multi-level fault propagation paths, thereby more accurately simulating the dynamic behavior of faults in the power grid. Overall, this formula combines the Laplace operator with multi-layer gradient changes to construct a dynamic model that can reflect the evolution of the power grid fault feature field over time. This dynamic model not only takes into account the spatial topological structure of the power grid, but also combines the multi-layer features extracted by the graph neural network, so that the system can describe the fault propagation path from multiple dimensions. Unlike traditional fault analysis methods, the present invention uses the feature extraction results of the multi-layer graph neural network as a basis, and accurately models the dynamic changes of the fault feature field through mathematical formulas, thereby providing strong support for fault identification and location. This formula uses mathematical gradients and Laplace operators to characterize the changing trend of power grid node features over time. The change of the fault feature field F(t) is jointly determined by the topological connection between nodes, the gradient of feature changes, and the spatial diffusion of fault propagation. Since each layer of the graph neural network captures node features in different ways, the system can obtain a global fault propagation model by comprehensively processing multi-layer features. In this way, when a fault occurs in the power grid, the system can accurately predict the fault propagation path, locate the fault source, and further analyze the impact range of the fault through the dynamic changes of the fault characteristic field.

[0060] Embodiment 6: Using the fault characteristic field, the fault probability distribution P(r, t) is calculated by the following formula:

[0061]

[0062] Where P(r, t) represents the fault probability distribution at location r and time t; Ω is the power grid domain.

[0063] Specifically, the first part of the formula It is the probability density function of the fault feature field F(t), which measures the change of the strength of the fault feature field F(t) over time in an exponential decay manner. The design of this part reflects the core influence of the fault feature field on the probability of fault occurrence. The norm of the feature field ∥F(t)∥ reflects the change of the state of the node in the power grid at a specific moment, and the exponential decay term is used to measure the significance of this change. In this way, the greater the intensity of the characteristic field, the greater the corresponding probability of failure. At the same time, the integral in the denominator represents the normalization factor over the entire power grid domain Ω, which ensures that the sum of the probability distribution in space is 1, that is, the probability of fault occurrence is a standardized distribution. The form of this probability density function reflects the relationship between the fault characteristic field and the power grid state: when the characteristic field intensity of a certain area is high, it means that the node state in the area has changed significantly, which may indicate the occurrence of a fault; while the area with low characteristic field intensity indicates that the node state is relatively stable and the probability of fault occurrence is low. Therefore, this part dynamically adjusts the probability of fault occurrence in different areas by changing the intensity of the characteristic field.

[0064] Next, the formula The term reflects the relationship between the topological structure of the power grid and the characteristics of the nodes. Here, ∥A∥ represents the norm of the adjacency matrix, which reflects the connection strength between nodes in the power grid. The topological structure of the power grid plays a vital role in the fault propagation process. Faults of certain nodes may propagate rapidly to adjacent nodes through power lines. Therefore, the norm of the adjacency matrix is ​​used to describe the overall situation of node connections in the power grid. And ∥D∥ represents the norm of the degree matrix, which is used to describe the number of connections of the node, that is, the interconnection between each node and other nodes. By combining the norm of the adjacency matrix with the norm of the degree matrix, the formula can balance the connection strength of different nodes in the power grid, thereby preventing some nodes with larger degrees from having too much influence on the overall probability distribution during the fault propagation process. In addition, the formula also contains ∥H l ∥, which represents the norm of the feature matrix of the last layer of the graph neural network. This term reflects the impact of the feature results extracted by the graph neural network on the fault probability distribution. The graph neural network extracts multiple layers of features and integrates the feature representation of each node in the power grid with the status information of its neighboring nodes. Therefore, H l The norm of is used to describe the overall change of the final feature matrix. By combining the norms of the adjacency matrix, degree matrix, and feature matrix, the formula can effectively capture the complex relationship between the grid topology and node characteristics, thereby more accurately calculating the probability of fault occurrence. Finally, the formula The nonlinear activation function tanh of the multi-layer graph neural network is introduced in this article to describe the nonlinear changes of the features of each layer of the graph neural network. The hyperbolic tangent function tanh(·) here is a common activation function that can limit the eigenvalues ​​between [-1,1] to avoid the situation where the eigenvalues ​​are too large or too small. At the same time, the nonlinear property of tanh can enhance the network's ability to recognize complex patterns, especially when dealing with complex systems such as power grids, where the features of multi-layer networks often show nonlinear changes. By computing the feature matrix H of each layer k By performing nonlinear activation, the system can capture the complex changing patterns during the propagation of power grid faults, thereby improving the accuracy of fault probability distribution.

[0065] Embodiment 7: Determine the fault area and fault boundary according to the fault probability distribution by the following formula:

[0066]

[0067] Among them, R fault is the fault area; B fault is the fault boundary; η th is the probability threshold; ξ th is the probability gradient threshold; n is the boundary normal vector; s is the boundary integral variable, which represents the position of each tiny line segment on the boundary of the fault area.

[0068] Specifically, the first part of the formula is used to determine the fault area R fault This process relies on two key conditions: the failure probability P(r, t) and the threshold setting of its gradient. According to the definition of the formula, the failure area R fault Include those that satisfy P(r, t)>η th The position r, that is, the failure probability exceeds the set threshold η th This threshold η th It is used to screen those areas with higher probability of failure, thereby effectively excluding areas with low probability of failure and reducing the possibility of false alarms. At the same time, the definition of the fault area also requires That is, the gradient value of the failure probability exceeds a set threshold ξ th This condition ensures that the system not only considers the absolute value of the failure probability, but also identifies the faulty area by analyzing the rate at which the failure probability changes over space. It reflects the rate of change of fault probability in the power grid space. If the gradient of a certain area is large, it means that the fault propagation characteristics of the area may be more severe. Therefore, combined with the gradient threshold ξ th , the system can more accurately identify the boundaries and internal areas of the fault-affected area. The introduction of this gradient threshold helps to improve the accuracy of fault area identification, especially during the propagation of power grid faults, where the fault effects of different nodes may spread at different rates, so relying solely on probability values ​​is not enough to accurately delineate the fault area. By adding gradient information, the system can identify areas where significant changes have occurred, further improving the accuracy of fault detection.

[0069] Fault Boundary B fault The determination is based on the fault area R fault In the formula, the fault boundary B fault The fault region boundary It is obtained by integrating the fault characteristic field F(t) on . Specifically, It means that the fault characteristic field F(t) is integrated along the boundary direction n on the boundary of the fault area. The integral variable s here represents the position of each tiny line segment on the boundary. By integrating the fault characteristic field on the boundary, the system can obtain the strength and direction information of the fault propagation on the boundary. The physical meaning of boundary integral is that it can reflect the characteristics of the fault at the boundary when it spreads from the inside of the area to the outside. Since the fault characteristic field F(t) represents the intensity and change of the fault state in the power grid, integrating the fault characteristic field on the boundary can effectively capture the impact intensity of the fault at the boundary. In this way, the system can determine the boundary of the fault area and understand the propagation trend of the fault on the boundary. This is of great significance for further analyzing the propagation path and affected range of the fault, especially in the case of complex power grid structure, the direction and intensity of fault propagation may be carried out along different paths. Therefore, through the boundary integral method, the system can accurately locate the boundary of fault propagation and ensure the accurate prediction of the fault propagation path. The whole formula demarcates the fault area by combining probability and gradient information, and determines the fault boundary by integrating the boundary of the fault area. This method can effectively handle the complexity of the power grid topology and accurately identify the impact range of the fault when a fault occurs. In traditional power grid fault detection methods, a single characteristic value or a simple threshold is usually relied on to delineate the fault area, which is prone to false alarms or missed alarms. In the present invention, by comprehensively utilizing the fault probability and its gradient information, the system can more accurately capture subtle changes in the fault propagation process, thereby ensuring that the fault area and boundary identification are more accurate. This method of fault area and boundary identification based on probability distribution and gradient provides strong support for the automated detection and location of power grid faults. Through the precise calculation of the fault probability distribution and the gradient change of the node state, the system can not only identify the location where the fault occurs, but also track the propagation path of the fault in the power grid. In addition, the introduction of boundary integrals provides the system with boundary information of fault propagation, helping operators to better understand the impact range and propagation trend of the fault. This refined fault detection and location method helps to improve the operational reliability of the power grid and reduce the economic losses and social impact caused by faults.

[0070] Embodiment 8: According to the fault area and the fault characteristic field, the fault propagation distribution is calculated by the following formula:

[0071]

[0072] Among them, T fault (t) is the fault propagation distribution at time t; τ is the time integral variable; is the tensor product.

[0073] Specifically, the fault propagation mode in the power grid is captured by performing tensor product operations on the fault area and fault feature field and combining time integration and exponential decay terms. The first part of the formula Represents the tensor product of the fault region and the fault feature field. The significance of the tensor product operation is to couple the information of the fault region with the dynamic changes of the fault feature field, thereby describing the propagation effect of the fault feature in a specific spatial region. Specifically, the fault region R fault (τ) represents the location of the fault in the power grid at time τ, while the fault feature field F(τ) reflects the state change of the area when the fault occurs. Through the form of tensor product, the strength, direction and other information of the fault area and the fault feature field can be combined to construct a complete fault propagation model. Next, the time integral term in the formula Represents the cumulative effect from the initial moment to the current moment t. The introduction of the integral operation is to consider the evolution of the fault at different moments, rather than just the characteristic changes at a certain moment. Since the propagation of faults in the power grid is often a dynamic process and may gradually spread to other areas over time, through time integration, the system can capture the propagation trajectory and evolution pattern of the fault. In the integration process, the time variable τ represents the historical moment traversed by the integral operation, which allows the system to accumulate the fault areas and characteristic changes of all past moments, thereby accurately describing the global dynamics of fault propagation.

[0074] The exponential decay term in the formula It is used to weigh the impact of feature changes at different time points on the current moment t. Here, X(t) and x(τ) represent the feature information at time t and historical time τ respectively, and the difference ∥X(t)-x(τ)∥ reflects the difference between the current feature and the past feature. Through exponential decay, the formula can dynamically adjust the contribution of the features at past moments to the current propagation distribution. When the feature difference is small, the value of the exponential term is large, indicating that the past features have a stronger impact on the current propagation distribution; conversely, when the feature difference is large, the value of the exponential term is small, indicating that the past features have a weaker impact on the current propagation. In this way, the system can focus on those historical feature changes that have a significant impact on the current moment according to the time change of the feature, so as to more accurately calculate the fault propagation distribution. In addition, the introduction of the time step Δt is to smooth the fault propagation process and avoid the instantaneous changes from causing drastic fluctuations in the propagation distribution. By reasonably selecting the time step, the system can effectively balance the cumulative effect of feature changes and the smoothing effect of exponential decay, thereby providing a more stable and accurate fault propagation distribution in the time dimension. Through this formula, the system can not only capture the propagation trajectory of the fault in the power grid, but also reflect the cumulative effect and dynamic changes of the fault through time integration. Compared with the traditional fault propagation analysis method, the present invention provides a more comprehensive fault propagation modeling method by combining mathematical tools such as tensor product, time integration and exponential decay. This method can simultaneously consider the spatial topological structure of the power grid, the spatiotemporal changes of node characteristics, and the dynamic diffusion of the fault area, so that the system can accurately predict the propagation path of the fault from a global perspective when facing complex power grid faults. Overall, the fault propagation distribution T fault The calculation formula of (t) integrates the spatiotemporal information of the fault area and the fault characteristic field, and introduces the dynamic adjustment of time integration and exponential decay, so that the system can fully describe the propagation process of the fault in the power grid. This design can not only improve the accuracy of fault detection, but also provide rich spatiotemporal information support for subsequent fault location and fault handling. Through this refined fault propagation analysis, the system can better understand the dynamic behavior of the fault, thereby helping power grid operators to take timely measures in the early stage of the fault, prevent the fault from further spreading, and reduce the impact of the fault on the power grid operation.

[0075] Embodiment 9: The fault location and the credibility of the calculated fault location are calculated by the following formula:

[0076]

[0077] Among them, L final is the fault location; C reliability is the fault location L final The credibility of; T is the total time.

[0078] Specifically, fault propagation in the power grid is a dynamic process, and the characteristic field continues to spread with the change of time and space. Therefore, it is difficult to fully characterize the fault propagation path and location by relying only on the fault characteristic information at a certain moment. In order to capture this dynamic behavior, the system transforms the fault propagation distribution T fault (t) and the fault probability distribution P(r, t) are combined to accumulate the propagation information in all time periods. This joint integration operation means that the system not only pays attention to the fault status at the current moment, but also takes into account the propagation trajectory at past moments, so as to more comprehensively reflect the overall trend of fault occurrence. The core logic of this fault location calculation is to find the spatial location that maximizes the cumulative effect, that is, the area where the fault is most likely to occur. Fault propagation distribution T fault (t) represents the diffusion path of the fault in the power grid. Over time, the fault may spread from the source to neighboring nodes. By combining the probability distribution P(r, t), the system can identify the probability of a failure at a certain node. Therefore, the time integration of this information can effectively take into account the spatiotemporal propagation process of the fault, and by maximizing these comprehensive features, find the spatial point that best matches the fault propagation and probability distribution, that is, the fault location L final .

[0079] This method based on the joint analysis of time and space characteristics can capture the dynamic changes of different regional characteristics in the power grid. In particular, when a power grid fails, the state changes between nodes are often time-space related. By integrating these changes, the system can identify the path of fault propagation and accurately locate its origin. Next, after determining the fault location, the system must also evaluate the reliability of the fault location. To this end, the present invention introduces the credibility of the fault location C reliability , by analyzing the difference between the fault area and the global features, the reliability of the positioning results is quantified. The change of the fault feature field F(t) is manifested in the space as the change of gradient and curl. Especially in the fault area, the local fluctuation of the feature field is often significant, reflecting the severity of the fault propagation. In order to quantify this local change, the system introduces curl to describe the directionality and complexity of the characteristic changes in the fault area. By incorporating the product of curl and gradient into the calculation, the system can capture the complex field changes inside the fault area, which means that when the characteristic field changes drastically in the fault area, the credibility of the fault location is higher. Conversely, if the field changes in the fault area are relatively gentle, the system may believe that the fault impact in this area is weak, thereby reducing the credibility of fault location. This quantification method based on curl and gradient can ensure that when the system locates the fault, it not only considers the size and location of the fault area, but also pays attention to the internal characteristic changes in the area, thereby further improving the accuracy of fault location.

[0080] At the same time, the global integral term ∫∫ in the denominator Ω ∥F(t)∥·P(r, t) represents the joint effect of the fault characteristic field strength and the fault probability in the global scope of the power grid. This global term is introduced to compare the local fault characteristic changes with the overall characteristics of the entire network. Through this global comparison, the system can measure the relative strength of the characteristic changes inside and outside the fault area, and further correct the credibility of the fault location based on this relative change. If the characteristic changes in the fault area are much greater than other areas in the power grid, then the system can judge that the positioning of the fault location is reliable; conversely, if the characteristic changes in the fault area are not significant, the credibility of the positioning may be low. The calculation of credibility is essentially a comparative analysis of local and global characteristic changes. By integrating the curl and gradient in the fault area, the system can effectively identify the characteristic fluctuations at the fault occurrence point, and at the same time ensure that these characteristic fluctuations are compared with the background of the entire power grid through global integration. This global and local comparative analysis enables the system to provide high-precision fault location results in complex power grid structures, and through credibility assessment, provides a quantitative evaluation standard to help operators judge the reliability of fault location.

[0081] As described above, the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features thereof may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A power grid topology fault location system based on graph neural network, characterized in that: The system includes: a power grid topology feature extraction unit, which is used to regard the power grid as a graph network, collect real-time data for each node of the power grid, record the real-time electrical measurement value of each node, and construct a feature matrix, wherein each element in the feature matrix represents the electrical characteristics of a node of the power grid at a certain time; a multi-layer graph neural network model unit, which is used to take the feature matrix as the input of the first layer, combine the adjacency matrix and the degree matrix of the power grid, and after multi-layer feature extraction of the preset multi-layer graph neural network model, output the result of the last layer as the feature extraction result; a fault identification and positioning unit, which is used to construct a fault propagation dynamics model based on the feature extraction result to obtain a fault feature field; calculate the fault probability distribution using the fault feature field; determine the fault area and the fault boundary according to the fault probability distribution; calculate the fault propagation distribution according to the fault area and the fault feature field; calculate the fault location and the credibility of the calculated fault location according to the fault propagation distribution; the feature matrix is: ； in, For the power grid in time The characteristic matrix when ; Represents the total number of nodes in the power grid; For Node In time The voltage amplitude of For Node In time The cosine value of the voltage phase angle; For Node In time The current amplitude; For Node In time The sine value of the current phase angle; For Node In time Active power of For Node In time Reactive power; is the reference value of power, used to normalize active power and reactive power; For Node In time The frequency deviation, that is, the actual frequency relative to the reference frequency The amount of deviation; For Node In time Impedance; is the base value of impedance; is the time step, which indicates the time difference between two times; is the decay parameter, which controls the rate at which the feature matrix decays over time and is a set value.

2. The grid topology fault location system based on graph neural network according to claim 1, characterized in that: Electrical measurements include: voltage magnitude, current magnitude, voltage phase angle, current phase angle, active power, reactive power, frequency deviation, and node impedance.

3. The grid topology fault location system based on graph neural network according to claim 2, characterized in that: The preset multi-layer graph neural network model has a total of Layer, the feature extraction result is obtained through the following formula: ； in, is the result of the last layer, as the result of feature extraction; is the degree matrix of the power grid; is the adjacency matrix of the power grid; For the The weight of the layer; is the activation function; For Hadamard; For the The result of the layer; for the input layer, .

4. The grid topology fault location system based on graph neural network according to claim 3, characterized in that: Based on the feature extraction results, the formula for constructing the fault propagation dynamics model is: ； in, is the subscript integer index; Represents the first Layer results; represents the gradient operator; is the Laplace operator; For time Fault characteristic field at time .

5. The grid topology fault location system based on graph neural network according to claim 4, characterized in that: Using the fault characteristic field, the fault probability distribution is calculated using the following formula: : ； in, Indicates at location Place, time The failure probability distribution when For the power grid domain.

6. The grid topology fault location system based on graph neural network according to claim 5, characterized in that: The fault area and fault boundary are determined according to the fault probability distribution through the following formula: ； ； in, is the fault area; is the fault boundary; is the probability threshold; is the probability gradient threshold; is the boundary normal vector; is the boundary integral variable, which represents the position of each tiny line segment on the boundary of the fault area.

7. The grid topology fault location system based on graph neural network according to claim 6, characterized in that: According to the fault area and fault characteristic field, the fault propagation distribution is calculated by the following formula: ； in, For time Fault propagation distribution when is the time-integrated variable; is the tensor product.

8. The grid topology fault location system based on graph neural network according to claim 7, characterized in that: The fault location and the credibility of the calculated fault location are calculated using the following formula: ； ； in, is the fault location; Fault location Credibility; For total time.

Images