Fault positioning method and system for real-model experiment distribution line

Abstract

The invention discloses a real-model experiment distribution line fault positioning method and system. The method comprises the following steps: acquiring a topological structure diagram of a target power distribution network in a real-model experiment; based on the topological structure diagram of the target power distribution network, combining a preset diagram input feature selection logic, and obtaining target power distribution network node and edge features as input of a power distribution network fault model to be established; according to the input of the to-be-established power distribution network fault model and the output of a preset power distribution network fault model, training the power distribution network fault model based on the improved graph neural network, the output of the preset power distribution network fault model being the probability of the line fault in the power distribution network; and according to the trained power distribution network fault model based on the improved graph neural network, real-type experiment power distribution line fault positioning is completed. The system and the method have good expansibility and adaptability, can be suitable for power distribution networks of different scales and structures, and have wide application prospects.

Description

Technical Field

[0001] The present invention relates to the technical field of real-type experimental distribution line fault location technology, and in particular to a real-type experimental distribution line fault location method and system. Background Art

[0002] In modern power systems, the distribution network is a key link connecting power stations and users, and its stability and reliability are crucial to the operation of the entire power system. However, due to the complex structure, variable topology and various fault types of the distribution network, traditional distribution network fault location methods often have problems such as low efficiency and low accuracy. With the development of artificial intelligence technology, the use of machine learning and deep learning methods to intelligently diagnose and locate distribution network faults has become a research hotspot.

[0003] In the field of distribution network fault location, traditional algorithms such as rule-based expert systems and signal processing-based methods can achieve fault location to a certain extent, but these methods usually rely on expert experience and are difficult to cope with the complex and changeable distribution network environment. In addition, these methods have low computational efficiency when processing large-scale data and are difficult to meet real-time requirements.

[0004] In recent years, deep learning technology has achieved remarkable results in image recognition, natural language processing and other fields. Its powerful feature extraction and pattern recognition capabilities provide new ideas for fault location in distribution networks. In particular, graph neural network models such as graph convolutional networks (GCN) and spatial graph convolutional networks (GraphSAGE) can directly operate on graph structured data, effectively capture the complex relationship between nodes in the distribution network, and provide new technical means for fault location in distribution networks.

[0005] However, the switch combination explosion problem and topological uncertainty of the distribution network have brought new challenges to fault location. In order to overcome these difficulties, researchers have begun to explore the distribution network load transfer model based on deep reinforcement learning. Deep reinforcement learning combines the feature extraction ability of deep learning and the decision-making ability of reinforcement learning, and can learn the optimal strategy through interaction with the environment. As an improved algorithm of deep reinforcement learning, the Dueling-DQN algorithm can more effectively evaluate the value of different actions by decomposing the Q-value function into a state value function and an advantage function, thereby improving the accuracy of decision-making.

[0006] In summary, this application aims to develop a new distribution network fault location algorithm by combining graph convolutional networks and deep reinforcement learning technology to improve the accuracy and robustness of fault location, while optimizing the load transfer strategy and improving the operating efficiency and reliability of the distribution network. Summary of the invention

[0007] In view of the above existing problems, the present invention is proposed.

[0008] Therefore, the present invention provides a real-type experimental distribution line fault location method and system to solve the problem of switch combination explosion in the distribution network and the challenge of fault location brought by topology uncertainty.

[0009] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0010] In the first aspect, the present invention provides a method for locating distribution line faults in a real-type experiment, comprising: obtaining a topological structure diagram of a target distribution network in a real-type experiment; according to the topological structure diagram of the target distribution network in the real-type experiment, combined with a preset graph input feature selection logic, obtaining the node and edge features of the target distribution network as the input of a distribution network fault model to be established; according to the input of the distribution network fault model to be established and the output of the preset distribution network fault model, training a distribution network fault model based on an improved graph neural network, the output of the preset distribution network fault model being the probability of a line failure in the distribution network; and completing the distribution line fault location in the real-type experiment according to the distribution network fault model based on the improved graph neural network after training.

[0011] As a preferred solution of the real-type experimental distribution line fault location method described in the present invention, the preset graph input feature selection logic includes:

[0012] The target distribution network node characteristics include six-dimensional three-phase voltage amplitude, phase angle positive sequence, negative sequence, zero sequence voltage amplitude, phase angle, two-dimensional neutral point voltage amplitude, phase angle, and six-dimensional three-phase injection / output active and reactive power;

[0013] The target distribution network edge features include six-dimensional three-phase line current amplitude, phase angle, positive sequence, negative sequence, zero sequence current amplitude, phase angle, six-dimensional three-phase line power amplitude, phase angle, and one-dimensional fault current alarm information;

[0014] Determine the model diagram input features according to the automation degree of the target distribution network, the type of the measuring device and the data category of the information source of the target distribution network;

[0015] Among them, all nodes are required to have the same input dimension, and all edges are required to have the same input dimension.

[0016] As a preferred solution of the real-type experimental distribution line fault location method described in the present invention, the distribution network fault model based on the improved graph neural network includes: an input layer, a hidden layer, and an output layer;

[0017] The input layer receives the model graph input features and outputs the processed graph input features to the hidden layer. The graph convolution algorithm used by the hidden layer is based on node feature data transmission and does not accept any edge feature data.

[0018] If the model graph input feature only selects the data of the target distribution network edge feature, the data of the target distribution network edge feature is directly input into the hidden layer;

[0019] If the model graph input feature contains data of the edge feature of the target distribution network, it must be formatted, the edge feature data must be converted into node feature data, and then output to the hidden layer;

[0020] The input layer includes merging the original edge features into node features through a single-layer neural network.

[0021] As a preferred solution of the real-type experimental distribution line fault location method of the present invention, the input layer includes merging the original edge features into node features through a single-layer neural network, including:

[0022] According to the dimension d of the edge feature vector e , construct a single-layer fully connected neural network and add a bias;

[0023] Input edge feature data E1 and E2 of each edge into the single-layer fully connected neural network to obtain output values ​​E'1 and E'2 of the same dimension;

[0024] Sum the neural network output vectors of all edges to get the sum vector E of the same dimension SUM ;

[0025] The node feature vector N V and the edge sum vector E SUM By Concat(N V , E SUM ) operations are merged into a new vector as output.

[0026] As a preferred solution of the real-type experimental distribution line fault location method of the present invention, the hidden layer includes:

[0027] Building a multi-layer graph convolutional neural network in the hidden layer;

[0028] The multi-layer graph convolutional neural network is constructed by stacking several single hidden layers, with node v i For example, the node features of the single hidden layer in the l+1 layer The aggregation formula is:

[0029]

[0030] Among them, N(v i ) is the set of neighbor nodes, is the node v in the lth layer j The node feature vector of For node v i The result vector after the neighbors are aggregated, aggregate l+1 (·) is the learnable aggregation function of the l+1th layer, concat(·) is the concatenation function, which concatenates two vectors horizontally in their original dimensions, σ(·) is the Relu activation function, and W is the learnable weight parameter matrix.

[0031] As a preferred solution of the real-type experimental distribution line fault location method of the present invention, the output layer includes:

[0032] The node feature data in the hidden layer are merged, and the edge features are output through the fully connected layer:

[0033] E 1,2 =σ(W 1,2 Concat(V1,V2)

[0034] Among them, V1 and V2 are the node feature vectors of the two end points of the edge (v1, v2), which are passed in from the hidden layer. 1,2 is the feature vector of the processed output edge (v1, v2), Concat(V1, V2) is the merging of the hidden layer feature vectors of nodes v1 and v2, W 1,2 is an m×n parameter matrix, σ(·) is the ReLU activation function, which outputs a feature vector with a length of m for each edge, and n represents the number of edges;

[0035] The number of outputs of the fully connected layer is the number of classifications of the distribution network fault identification task, that is, the number of lines in the distribution network. Each value in the output vector represents the probability of a corresponding line failure, and the sum of all values ​​is 1.

[0036] As a preferred solution of the real-type experimental distribution line fault location method described in the present invention, the distribution network fault model based on the improved graph neural network also includes:

[0037] The loss function calculation formula in the distribution network fault model based on the improved graph neural network is:

[0038]

[0039] Among them, T is the label vector, Y is the predicted output vector, and y a is the probability of the ath fault predicted by the model, w aThe weight factor corresponding to the ath fault type is W1=[w1,w2,...,w n ] is the weight vector, A represents the total number of faults, t a represents the true label.

[0040] In a second aspect, the present invention provides a real-type experimental distribution line fault location system, comprising:

[0041] The data acquisition module is used to obtain the topological structure diagram of the target distribution network in the real experiment;

[0042] The input acquisition module is used to obtain the node and edge features of the target distribution network as the input of the distribution network fault model to be established according to the topological structure diagram of the target distribution network in the real experiment and the preset diagram input feature selection logic;

[0043] A model training module, used to train a distribution network fault model based on an improved graph neural network according to the input of the distribution network fault model to be established and the output of a preset distribution network fault model, wherein the output of the preset distribution network fault model is the probability of line failure in the distribution network;

[0044] The fault location module is used to complete the fault location of the real experimental distribution line according to the distribution network fault model based on the improved graph neural network after training.

[0045] In a third aspect, the present invention provides a computing device, comprising: a memory and a processor;

[0046] The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions. When the computer-executable instructions are executed by the processor, the steps of the method are implemented.

[0047] In a fourth aspect, the present invention provides a computer-readable storage medium storing computer-executable instructions, which implement the steps of the method when executed by a processor.

[0048] Compared with the prior art, the beneficial effects of the present invention are as follows: the present invention proposes a real-type experimental distribution line fault location method and system, which can obtain the distribution network node and edge features more accurately by obtaining the topological structure diagram of the target distribution network and combining the preset graph input feature selection logic, and provide richer information for the establishment of the fault model. The improved graph neural network is used to model the distribution network, which can effectively capture the complex topological relationships and the interactions between the node and edge features in the distribution network, thereby improving the accuracy of fault location. Real-time monitoring of the operating status of the distribution network and rapid fault location using the trained fault model can detect and handle faults in a timely manner, reduce power outage time, and improve the operating reliability of the distribution network. The system and method of the present invention have good scalability and adaptability, can be applied to distribution networks of different scales and structures, and have broad application prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0050] Figure 1 is an overall flow chart of the method described in the embodiment of the present invention;

[0051] Figure 2 A schematic diagram of a distribution network structure of the method according to an embodiment of the present invention;

[0052] Figure 3 A schematic diagram of a typical power distribution network structure of the method described in an embodiment of the present invention;

[0053] Figure 4 A schematic diagram of an input layer of the method according to an embodiment of the present invention;

[0054] Figure 5 A schematic diagram of an LSTM aggregator of the method described in an embodiment of the present invention;

[0055] Figure 6 A schematic diagram of graph convolutional neural network feature transfer of the method described in an embodiment of the present invention;

[0056] Figure 7 This is a schematic diagram of the output layer structure of the method according to an embodiment of the present invention;

[0057] Figure 8 A schematic diagram of an improved IEEE33 node power distribution system of the method according to an embodiment of the present invention;

[0058] Fig. 9A power flow diagram of an IEEE33 node according to the method of an embodiment of the present invention;

[0059] Fig.10 A schematic diagram for comparing different input feature results of the method according to an embodiment of the present invention;

[0060] Fig.11 This is a schematic diagram comparing the accuracy of different algorithms of the method described in the embodiment of the present invention;

[0061] Fig.12 A schematic diagram for comparing different aggregation functions of the method according to an embodiment of the present invention;

[0062] Fig.13 The figure is a schematic diagram for comparing the Loss of different algorithms of the method described in the embodiment of the present invention. DETAILED DESCRIPTION

[0063] In order to make the above-mentioned purposes, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the drawings of the specification. Obviously, the described embodiments are part of the embodiments of the present invention, but not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary persons in the art without creative work should fall within the scope of protection of the present invention.

[0064] Example 1

[0065] Reference Figure 1 to Figure 7 As an embodiment of the present invention, a real-type experimental distribution line fault location method and system are provided;

[0066] Before describing the embodiments of the present application in detail, some related concepts are first explained for the sake of clarity.

[0067] Real-type experiment: Real-type experiment usually refers to the experiment conducted under the actual use conditions or environment, which can reflect the actual operation as much as possible. In this context, "real-type experiment" means testing the fault location technology in a simulated real distribution network environment to ensure that it can work effectively in the actual distribution network.

[0068] Multi-layer graph convolutional neural network: Multi-layer graph convolutional neural network (Multi-layer Graph Convolutional Networks) is a deep learning model specially designed to process graph structured data. In traditional convolutional neural networks, data is usually in a grid-like shape, such as image data, but in graph structured data, the connections between data points (or nodes) are irregular, which is manifested in the distribution network as nodes (such as power nodes, load nodes) connected to each other through lines (edges). Graph convolutional neural networks take advantage of this feature and extract local and global patterns or features in the graph by transferring node features between different layers of the network.

[0069] Compared with the transmission network, the distribution network has many differences. The voltage level of the distribution network is lower and its power supply range is relatively small, but because it is directly connected to the user, the distribution network can be said to be the first line of customer service and the link between the transmission side and the user side. As the hub between lines of different voltage levels and the "last mile" of the power supply link, the distribution network provides stable power supply services, reduces line tripping and power outages, improves the reliability and stability of the distribution network, and improves the overall level of user satisfaction with electricity consumption.

[0070] At present, there are three main types of distribution network wiring as shown in the figure, including radial, tree and ring. Figure 2 As shown, the real-type experimental field of this application can cover Figure 2 All common topologies.

[0071] The advantages of radial distribution network are short construction period, low investment cost, simple and easy wiring form, but the disadvantage is that it cannot fully guarantee the reliability of power supply; the advantages of tree-type distribution network are long power supply radius and simple and easy wiring form, but the reliability is also low, and the disconnection of the trunk fault will cause all downstream loads to lose power; the ring-type distribution network, based on the radial distribution network, realizes double-end power supply by adding an open-loop operation of a middle tie switch, and the two ends serve as backup for each other, thereby greatly improving its power supply reliability. Among the above three wiring methods, the ring-type distribution network has the highest power supply reliability and is currently the most widely used wiring method in my country's distribution network. In order to avoid the generation of network circulation and through power, the tie switch of the ring-type distribution network is in a disconnected state under normal operation. If one end fails and the power is cut off, the tie switch automatically closes, and the non-fault area receives continuous power supply from the other end, which fully improves the reliability of power supply.

[0072] The probability of failure of distribution network is higher than that of transmission network. About 80% to 90% of power outages are caused by distribution network failures. The main factors that lead to distribution network failures are as follows:

[0073] 1) The power supply lines are long, the grid is complex and continues to grow, the early investment is small, the types and quantity of distribution network equipment are large, but the level of distribution equipment is uneven, the overall quality is not high, the operation and maintenance means are insufficient, and potential fault hazards are difficult to detect.

[0074] 2) The distribution network is often close to the city center or residential areas, so the environment is relatively complex. There may be distribution network failures caused by construction or vehicle accidents, as well as human vandalism and illegal activities such as theft of power equipment.

[0075] 3) The severe weather in recent years is also the main cause of distribution network failures. Rising temperatures accelerate equipment aging and pose a greater threat to the working environment of distribution equipment.

[0076] 4) With the rise of new energy sources such as wind and solar power, a large number of distributed power sources are connected to the distribution network. The distribution network containing distributed power sources will provide more fault current to the fault point in the event of a fault, which may easily cause the fault range to expand further. In addition, it is difficult to determine the location of the fault in a multi-source distribution network, which brings difficulties to isolating the fault and reducing the power outage area.

[0077] There may be unbalanced loads in the distribution network, but power supply companies usually distribute the three-phase loads evenly. Therefore, in normal operation, the zero-sequence and negative-sequence components in the distribution network can usually be ignored. One of the most common faults in the distribution network is the single-phase grounding fault, which usually accounts for 80% of the distribution network faults. The zero-sequence voltage is the highest at the single-phase grounding fault, and the fault point current is the capacitive current of the three-phase line to the ground. The increase in the non-fault phase voltage may threaten the insulation, and then break down the insulation and develop into a multi-phase, multi-point fault. Therefore, in practice, once a single-phase grounding fault occurs, the location of the fault must be found as soon as possible so that it can be controlled and isolated in time to prevent further deterioration and cause greater losses to the entire system.

[0078] There is no zero-sequence component in a two-phase short circuit. The negative-sequence voltage is the highest at the short-circuit point, which is equal to the positive-sequence voltage in magnitude and opposite in direction. A two-phase short circuit to grounding produces zero-sequence voltage and zero-sequence current, which reach the maximum value at the fault point. The consequences of a three-phase short circuit are the most serious. The three-phase current increases and the voltage decreases, but the three-phase symmetry is still maintained, and there is no zero-sequence current or zero-sequence voltage.

[0079] Supervisory Control And Data Acquisition (SCADA) is a computer-based distributed control system (DCS) and power automation monitoring system; it has a wide range of applications and can be applied to data acquisition and monitoring control and process control in industrial fields such as electricity, metallurgy, and petroleum. In the power system, the SCADA system is the most widely used and the technology is the most mature. It occupies an important position in the telecontrol system and can monitor and control the operating equipment on site to realize various functions such as data acquisition, equipment control, measurement, parameter adjustment, and various signal alarms, which are known as the "four remote" functions. The feeder terminal unit (FTU) is an important part of it. The dispatch center only needs to remotely control the FTU and other terminal devices through the SCADA system to realize various complex functions and operations, which plays a very important role in the construction of today's distribution automation system.

[0080] The feeder terminal unit FTU is a switch monitoring device installed next to the feeder switch. These feeder switches refer to outdoor pole switches, such as circuit breakers, load switches, and section switches on 10kV lines. The pole switches are monitored to complete telemetry, remote control, remote communication, and fault detection functions, and communicate with the distribution automation master station to provide the distribution system operation status and various parameters, that is, the information required for monitoring and control, including switch status, power parameters, phase-to-phase faults, ground faults, and parameters at the time of faults, and execute the commands issued by the distribution master station to adjust and control the distribution equipment, thereby realizing automation functions such as rapid fault location, fault isolation, and rapid restoration of power supply to non-fault areas. It has been widely used in distribution network fault location technology.

[0081] When a line fails, the FTU can record and upload the line data before and after the failure in the first time. The SCADA system quickly locates the section where the failure occurs by analyzing and calculating the data uploaded by the FTU, and issues corresponding instructions to the corresponding FTU to control the opening and closing status of the switch to isolate the fault area and restore power supply to the non-fault area. Therefore, the transmission of fault information is of great significance to the efficiency of fault handling, shortening the power outage time, reducing the power outage area and improving the reliability of power supply. The real-type experimental field of this application simulates the actual distribution network, equipped with a SCADA system similar to the actual distribution network and its FTU, DTU and other terminal equipment, but for experimental purposes, it is different from the SCADA system of the actual distribution network, and special consideration is given to the intelligent integration with the digital simulation system.

[0082] In related technologies, there are some challenges, such as the complexity and dynamic characteristics of distribution networks make fault location difficult. Traditional fault detection methods rely on fixed rules and thresholds, which are often not flexible and accurate enough when facing changes in distribution network structure or abnormal situations. In addition, the noise and uncertainty factors that may exist in the distribution network will also affect the performance of fault detection.

[0083] The present application provides a method that can effectively solve the above-mentioned problems. Next, multiple embodiments will be combined to explain in detail how to implement the real-type experimental distribution line fault location method; Figure 1 The present invention is a method flow chart of a real-type experimental distribution line fault location method and system, which specifically includes the following steps:

[0084] S100: Obtain the topological structure diagram of the target distribution network in the real experiment;

[0085] Specifically, the distribution network is a graph with obvious topological characteristics, consisting of numerous load nodes and segment switches connecting the nodes. The typical distribution network structure is as follows: Figure 3 As shown in the figure, the distribution network has an active node SG, a distributed power source DG, and 6 load nodes, among which load node 4 and DG are located at the same node, and the nodes are connected by various segmented switch FTU units. According to the definition of graph in graph neural network, the graph is composed of vertices V and edges E connecting the vertices. Therefore, the above figure is abstracted into a graph composed of points and edges, and in the distribution network, both the power nodes and the load nodes have a relationship of mutual influence and mutual restriction.

[0086] Therefore, when the distribution network is abstracted as a graph, it should be an undirected graph. The above distribution network is abstracted to a graph G, expressed as G = (V, E), where the node set V has a total of 7 nodes, consisting of source nodes and 6 load nodes, and node 4 is a load node with distributed power supply. The edge set E has 6 edges, all of which are undirected edges formed by the switches of the segmented switch FTU unit. When constructing a graph neural network model for training, its input is mainly divided into two parts. The first part is the graph G containing only topological structure information, and the second part is the graph signal input on the nodes and edges corresponding to the graph structure. These graph signal input information represents the state characteristics of the nodes and edges. For the distribution network, there are many node and edge features that can be selected. The feature information that is conducive to the graph neural network to determine the fault location should be selected as the input of the graph signal according to actual needs.

[0087] S200: Based on the topological structure diagram of the target distribution network and in combination with the preset graph input feature selection logic, the node and edge features of the target distribution network are obtained as inputs of the distribution network fault model to be established;

[0088] In the embodiment of the present application, the preset graph input feature selection logic includes:

[0089] The target distribution network node characteristics include six-dimensional three-phase voltage amplitude, phase angle positive sequence, negative sequence, zero sequence voltage amplitude, phase angle, two-dimensional neutral point voltage amplitude, phase angle, and six-dimensional three-phase injection / output active and reactive power;

[0090] The target distribution network edge features include six-dimensional three-phase line current amplitude, phase angle, positive sequence, negative sequence, zero sequence current amplitude, phase angle, six-dimensional three-phase line power amplitude, phase angle, and one-dimensional fault current alarm information;

[0091] and determining the model diagram input characteristics according to the automation degree of the target distribution network, the type of the measuring device, and the data category of the information source of the target distribution network;

[0092] All nodes are required to have the same input dimension, and all edges are required to have the same input dimension;

[0093] It should be noted that for all types of distribution network faults, the positive sequence voltage gradually decreases from the power source point to the short circuit point; for asymmetric faults, the fault point is often the point where the zero sequence and negative sequence voltage amplitudes in the distribution network are the largest. In real situations, the probability of a three-phase short circuit is extremely small, so using the zero sequence, negative sequence voltage, current, etc. at each node of the distribution network will be very helpful in locating the location of the distribution network fault. Therefore, the three-sequence voltage and current obtained by decomposition are added to the table.

[0094] It should be noted that the nodes V of the distribution network are mainly load nodes and distributed power supply nodes, and the node features are mainly the fault electrical quantity data uploaded by the terminal, such as the active / reactive power information generated / consumed by the three phases, a total of 6 node dimensions; the voltage amplitude, phase angle and other information of the three phases, a total of 6 node dimensions; and the neutral point potential information, 2 node dimensions of amplitude and phase angle. However, since some nodes are connected in an angular manner, there may not actually be a neutral point, and the neutral point potential information cannot be obtained, and the measuring device FTU may not have the function of measuring the neutral point voltage. Further processing of the original data can obtain the positive sequence, negative sequence, and zero sequence node voltage components, a total of 6 node dimensions. Some distribution networks may not be equipped with GPS devices to determine the precise time, so it may not be possible to synchronously measure the phase angle of electrical quantities such as voltage and current. At this time, only the amplitude can be used as the input feature, and the corresponding feature dimension becomes 3. The decomposition of sequence components does not require the calculation of synchronous phases. It only needs to obtain the voltage and phase data of the three phases of a single measuring device at a certain moment. The main amplitude of the sequence component is of great reference significance for fault location, and no phase angle data is required. There are three dimensions of node data in total. Most faults in the distribution network are asymmetric faults. Selecting sequence components as node inputs will have a better fault location effect than three-phase voltage.

[0095] It should be noted that the edge E of the distribution network is mainly composed of distribution lines and switches. The inherent parameters of the line include data such as three-phase line impedance and line length. However, since they are fixed values ​​and will not change with the state of the distribution network, they are not very meaningful as inputs to the graph neural network. The electrical quantity information includes the three-phase current data flowing through the FTU terminal and the three-phase power data passing through. The edge feature dimension is 6, which is the same as the node voltage sequence component. The current sequence component can also be used as the edge input. When a fault occurs in the distribution network, a large fault current often passes through. In a distribution network with a low degree of automation, the FTU usually only uploads fault overcurrent alarm information. In a single radial distribution network, there are: state 0, normal; state 1, passing the fault overcurrent in the positive direction. In a distribution network with distributed power sources, there will also be a state of -1, representing a fault overcurrent in the reverse direction, and the dimension of the fault alarm information is 1.

[0096] It should be noted that, according to the degree of automation of the distribution network and the type of measuring device, when determining the model input characteristics, it should be considered what categories of data the information source of the actual distribution network has, and appropriate input characteristics should be selected as model input according to local conditions. The final model node and edge data dimensions should be the sum of the selected node and edge data dimensions.

[0097] For example, assume that a distribution network has the following characteristics:

[0098] ①FTU can measure the three-phase voltage amplitude and phase angle;

[0099] ②FTU can report fault current alarm information;

[0100] ③The system can measure positive sequence, negative sequence, and zero sequence voltage and current;

[0101] ④No neutral point voltage measurement;

[0102] ⑤ There is no GPS synchronization phase angle, so only the voltage amplitude is used.

[0103] Based on this information, the following features can be selected as input:

[0104] For nodes: three-phase voltage amplitude (3D) + positive sequence, negative sequence, zero sequence voltage amplitude (3D), a total of 6 dimensions; for edges: fault current alarm information (1D). Assuming that each edge reports fault current alarm information, the total input feature dimension is the node feature dimension plus the edge feature dimension. If there are 7 nodes and 6 edges, the total node feature dimension is 7×6=42 dimensions, and the edge feature dimension is 6×1=6 dimensions. The final input data dimension is the node feature dimension plus the edge feature dimension, that is, 42+6=48 dimensions.

[0105] It should be noted that the benefit of the above step S200 is that it can flexibly select and combine graph input features according to the characteristics of the actual distribution network, thereby improving the accuracy and efficiency of the graph neural network model in fault location of the distribution network. In this way, the model can better adapt to the automation level and measurement device type of different distribution networks, ensuring the best fault detection and location effect in all situations.

[0106] In addition, by presetting the graph input feature selection logic, we can ensure the consistency and standardization of input data when building the graph neural network model, thereby avoiding model training problems caused by inconsistent data. The input dimensions of all nodes and edges are consistent, which helps to simplify the model structure design and improve computing efficiency.

[0107] In practical applications, choosing the right graph input features is crucial to the performance of the model. For example, for node features, choosing the three-phase voltage amplitude and sequence component voltage amplitude can effectively reflect the electrical state of the node, while the fault current alarm information provides a direct fault indication for the edge feature. This feature combination can not only help the model quickly locate the fault point, but also reduce false positives and false negatives to a certain extent.

[0108] S300: training a distribution network fault model based on an improved graph neural network according to an input of a distribution network fault model to be established and an output of a preset distribution network fault model, wherein the output of the preset distribution network fault model is a probability of a line fault in the distribution network;

[0109] In an embodiment of the present application, a distribution network fault model based on an improved graph neural network includes: an input layer, a hidden layer, and an output layer;

[0110] The input layer receives the model graph input features and outputs the processed graph input features to the hidden layer. The graph convolution algorithm used by the hidden layer is based on the node feature data transmission and does not accept any edge feature data.

[0111] If the model graph input feature only selects the data of the target distribution network edge feature, the data of the target distribution network edge feature is directly input into the hidden layer;

[0112] If the model graph input features contain data of the target distribution network edge features, they must be formatted, the edge feature data must be converted into node feature data, and then output to the hidden layer;

[0113] The input layer includes merging the original edge features into node features through a single-layer neural network.

[0114] It should be noted that the input layer receives the input feature vector and outputs it to the hidden layer after processing. However, the graph convolution algorithm used by the hidden layer is based on the transmission of node feature data and does not accept any edge feature data. If only part of the distribution network node data is selected in the input data, the input layer can be omitted and the node data can be directly input into the hidden layer. If the input data includes edge feature data such as edge fault current alarm information data, three-phase current phase angle data, etc., it must be formatted, and the edge feature data must be converted into node feature data before being output to the hidden layer.

[0115] Furthermore, the input layer includes merging the original edge features into node features through a single-layer neural network, including:

[0116] Inspired by the idea that fully connected neural networks can transform input and output data in any dimension, the following Figure 4 As shown in the input layer, assume that node v has two adjacent edges, according to the dimension d of the edge feature vector e , construct a single-layer fully connected neural network, each neural network is d e input, d e outputs and add bias;

[0117] Input the edge feature data E1 and E2 of each edge into a single-layer fully connected neural network to obtain output values ​​E'1 and E'2 of the same dimension;

[0118] Sum the neural network output vectors of all edges to get the sum vector E of the same dimension SUM ;

[0119] The node feature vector N V and the edge sum vector E SUM By Concat(N V , E SUM ) operations are merged into a new vector as output.

[0120] Furthermore, the hidden layer includes:

[0121] Build a multi-layer graph convolutional neural network in the hidden layer;

[0122] It should be noted that the input layer processes the original graph feature data and merges it into a new node feature vector and passes it to the hidden layer. A multi-layer graph convolutional neural network is built in the hidden layer to realize the layer-by-layer transmission of node features. During the transmission process, the graph convolutional network gradually extracts local abstract features.

[0123] It should be noted that the frequency domain convolution algorithm Spectral GCN treats all nodes in the graph as a whole to calculate the overall distribution of its frequency domain signals, and its learnable weight parameters are learned together as a whole structure, but the flexibility in controlling the number of layers of information aggregation and distinguishing node types is relatively poor. The spatial domain convolution algorithm Spatial GCN structure has a relatively high design flexibility, which enhances the adaptability of graph convolutional neural networks to graph data. This more flexible network construction method is mostly used in graph convolutional network programming tasks, and the spatial domain convolution method basically includes the existing time domain convolution method and improves it. Therefore, this application also adopts a basic framework construction model based on spatial domain convolutional neural networks.

[0124] It should be noted that the neighbor sampling and aggregation functions proposed by the GraphSAGE algorithm have achieved good results in practical applications. For large-scale graphs, neighbor sampling solves the problem of training when there are too many neighbor nodes. For example, in complex social relationships, the average node degree may be thousands. However, most nodes in the distribution network are intermediate nodes of branch groups, and their degree values ​​are all 2. The degree of individual nodes is at most 3 to 5. Sampling a few neighbors may cause data duplication or loss. Therefore, this application does not recommend the application of neighbor sampling in distribution network-related issues. The aggregation function defines the way neighbor nodes affect the target node, which can be compared to the convolution kernel in the frequency domain graph convolution, which is used to extract local information of the target node and enhance the generalization ability of the model.

[0125] The multi-layer graph convolutional neural network is constructed by stacking several single hidden layers, with node v i For example, the node features of a single hidden layer at layer l+1 The aggregation formula is:

[0126]

[0127] Among them, N(v i ) is the set of neighbor nodes, is the node v in the lth layer j The node feature vector of For node v i The result vector after the neighbors are aggregated, aggregate l+1 (·) is the learnable aggregation function of the l+1th layer, concat(·) is the concatenation function, which concatenates two vectors horizontally in their original dimensions, σ(·) is the Relu activation function, and W is the learnable weight parameter matrix.

[0128] It should be noted that the three aggregation functions proposed by GraphSAGE are mean-based aggregation (Meanaggregator), long short-term memory network aggregation (LSTM aggregator) and pooling aggregation (Poolingaggregator).

[0129] The mean aggregation function takes the mean of neighbor nodes as the aggregation result. Its calculation formula is as follows:

[0130]

[0131] The mean aggregation function averages the feature values ​​of itself and its neighboring nodes and then passes them through the fully connected layer. It is a linear approximation of the frequency domain convolution GCN method. It has no algorithmic merging operation and physically forms an effect similar to a skip connection, which greatly improves the model.

[0132] LSTM aggregator such as Figure 5 As shown in the figure, compared with the mean aggregator, the LSTM aggregator has stronger expressive power and can learn more input distribution characteristics. The relevant nodes input the complex aggregator long short-term memory network and use the output as the aggregation result.

[0133] The circled areas a, b, and c in the figure represent the forget gate, update gate, output gate, and sequence state C respectively. t The information channel throughout the sequence, h t Represents the hidden state value transmitted at the previous moment; x t Represents the new input vector.

[0134] Each time LSTM completes the input x t , after the operation, a new sequence state and hidden state value are generated and input to the next moment. The calculation process is:

[0135] i t =σ(W i ·[h t-1 ,x t ]+b i )

[0136]

[0137] o t =σ(W o [h t-1 ,x t ]+b o )

[0138] h t =o t ⊙tanh(C t )

[0139] It should be noted that LSTM is usually used in time series tasks. Its input is a sequence in chronological order and does not have the symmetry of the aggregation function. Therefore, it is necessary to randomly shuffle the order of the neighbor nodes first, and then use the eigenvalues ​​h of the neighbor nodes as the input of LSTM in turn. The output of the final LSTM network is used as the aggregated representation of the eigenvalues ​​of the neighbor nodes. Its function is the same as the aggregation function in formula A, and it needs to pass through formula B before output.

[0140] Pooling aggregator uses the pooling operation in CNN to perform aggregation, taking the maximum value of neighbor nodes passing through the fully connected layer as the aggregation result. Its calculation formula is as follows:

[0141]

[0142] Aggregation After that, they still need to be merged through formula A. This method effectively captures the prominent feature performance on the adjacent nodes, and its computational complexity is between that of the MEAN and LSTM algorithms.

[0143] The single hidden layer is constructed by formula A and B. By stacking multiple layers of graph convolutional neural network, the node feature vector is transmitted layer by layer, such as Figure 6 As shown: Figure 6 The node feature vector value of the node v1 in the third layer is obtained from the first-order subgraph of the node v1 in the second layer, and the subgraph of the second layer is obtained from the second-order subgraph of the node v1 in the first layer. Therefore, the higher the number of layers, the larger its representation range and the more nodes are affected by the input. At the same time, because it is a comprehensive reflection of a large range of features, the local abstraction ability of the graph is stronger. However, the more layers the better. Stacking too many graph convolution layers may cause signal drowning, and too large a representation range will cause the graph convolutional neural network node representation to cover the entire graph, similar to ordinary neural networks, and lose the local representation characteristics of the graph.

[0144] In the embodiments of the present application, one of the tasks in which the current graph convolutional neural network method is widely used is node classification and regression. Each node in the training, validation and test sets in the graph has a category label assigned from a set of predefined categories. The purpose of distribution network fault location is to find the faulty line, which requires that the graph convolutional neural network should output the label of the faulty line when performing a single fault location task in the distribution network. Therefore, the distribution network fault location task is essentially an edge classification task rather than a node classification task. The GraphSAGE algorithm used for the transfer of hidden layers is a layer-by-layer transfer of node feature vectors, and its final output is also the feature vector of the node, so the present application designs a classification model for converting from nodes to edges, such as Figure 7 As shown, the output layer includes:

[0145] E 1,2 =σ(W 1,2 Concat(V1,V2)

[0146] Among them, V1 and V2 are the node feature vectors of the two end points of the edge (v1, v2), which are passed in from the hidden layer. 1,2 is the feature vector of the processed output edge (v1, v2), Concat(V1, V2) is the merging of the hidden layer feature vectors of nodes v1 and v2, W 1,2 is an m×n parameter matrix, σ(·) is the ReLU activation function, which outputs a feature vector with a length of m for each edge, and n represents the number of edges;

[0147] In the embodiment of the present application, after all edge feature vectors of the graph are obtained by the above algorithm, all edge feature vectors can be fully connected again to finally extract the output vector, and softmax is selected as the activation function. The calculation formula is:

[0148]

[0149] Among them, O(E) is the set of eigenvectors of all edges in the graph, W i，j is the parameter matrix corresponding to each feature vector, and y is the output vector of the classification task. The softmax calculation formula is:

[0150]

[0151] Where X represents the input vector, xi represents each element in the vector X, and yi represents the corresponding element in the output vector y;

[0152] The number of outputs of the fully connected layer is the number of classifications of the distribution network fault identification task, that is, the number of lines in the distribution network. Each value in the output vector represents the probability of a corresponding line failure, and the sum of all values ​​is 1.

[0153] It should be noted that the benefit of building a model according to the above steps is that it can fully utilize the advantages of graph convolutional neural networks in processing graph structure data, effectively fuse node features and edge features, thereby improving the accuracy and efficiency of fault location. In addition, by stacking multiple layers of graph convolutional neural networks, higher-level feature representations can be extracted layer by layer, allowing the model to capture more complex graph structure information.

[0154] In an embodiment of the present application, it is assumed that there is a distribution network with multiple possible fault points. Different from the traditional one-hot encoding, an extended multi-label encoding method will be adopted to allow a sample (ie, a line) to correspond to multiple fault types. For example, if there are multiple faults or a complex situation of faults, it can be represented by a vector, and each element in the vector corresponds to a possible fault type. If a certain fault type exists, the value of the position is 1, otherwise it is 0. This allows the model to detect multiple faults at the same time. Assuming there are n possible fault types, the label T is a vector of length n T = [t1, t2, ..., t n ], where t i ∈0,1 indicates whether the i-th fault occurs.

[0155] In the process of supervised learning neural network training, the error index between the neural network and the ideal accurate model is called the loss function. There are also many ways to measure the error of the ideal model. In theory, the loss function can use any function, but generally uses the root mean square error (RMSE) and cross entropy error (CrossEntropy Loss). The root mean square error is often used in regression tasks to measure the gap between the predicted value and the true value. The cross entropy error is often used in classification tasks to evaluate the error between the classification distribution probability and the true value. This application proposes an extended multi-label encoding method based on the proposed distribution network fault model based on the improved graph neural network. The calculation formula of the loss function is improved as follows:

[0156] The loss function calculation formula in the distribution network fault model based on the improved graph neural network is:

[0157]

[0158] Among them, T is the label vector, Y is the predicted output vector, and y a is the probability of the ath fault predicted by the model, w a The weight factor corresponding to the ath fault type is W1=[w1,w2,...,w n ] is the weight vector, A represents the total number of faults, t a represents the true label.

[0159] When t a =1, that is, when the ath fault does occur, the model should predict a high probability y a , the loss function penalizes cases with low predicted probabilities.

[0160] On the contrary, when t a =0, that is, when the ath fault does not occur, the model should predict a low probability y a, in which case the loss function penalizes those cases with higher predicted probabilities.

[0161] This loss function encourages the model to output the probability y a With the true label t a When the prediction matches the actual situation, the loss decreases; otherwise, the loss increases.

[0162] In this way, the model can continuously adjust its own parameters during the training process, making the prediction results closer to the actual situation, thereby improving the accuracy of fault location.

[0163] In practical applications, the weight factor w a The settings can be based on the frequency of the fault type, for example:

[0164] If fault type i is common, then w a Can be set smaller since this type of fault is easier to detect.

[0165] If failure type a is rare but has serious consequences, then w a Can be set larger to ensure that the model pays more attention to this type of failure during training.

[0166] It should be noted that if there are some very rare but serious fault types in the distribution network, a new loss function can also be used to solve it. The new loss function takes into account different very rare but serious fault types. The reason for proposing a new loss function is to prevent no less than two very rare but serious faults from occurring during the implementation process, namely:

[0167] When there are two or more very rare but serious failures, a new loss function is used. The new loss function is expressed as follows:

[0168]

[0169] Where L represents the total loss; n represents the total number of failures; represents the weight factor of the a-th very rare but serious fault type. For rare but serious fault types, the weight factor Should be larger; Represents the weight factor of the ath common fault type. For common fault types, the weight factor Can be smaller; a Represents the elements in the label vector, indicating whether the ath fault occurs (y a =1 means it happens, y a =0 indicates no occurrence); p a represents the probability of the ath fault predicted by the model; log(p a) and log(1-p a ) are the logarithm of the predicted probability and the logarithm of the probability of non-event occurrence, respectively. With this design, a greater weight is assigned to very rare but serious failure types. In order to pay more attention to these fault types during training. The weights of common fault types Small, because these faults appear more frequently in the training set and the model is easy to learn. Such a loss function helps improve the model's ability to identify key fault types and overall generalization ability.

[0170] In the embodiments of the present application, very rare but serious fault types include multiple phase-to-phase short circuits, multi-point ground faults, large-scale transient overvoltages, reverse power flow of distributed power sources, and faults caused by extreme weather conditions;

[0171] Specifically, multiple phase short circuits: This type of fault involves direct contact between two or more conductors, resulting in huge short-circuit currents. Although this type of fault is less common, it can cause equipment damage or even a large-scale collapse of the power grid. The consequences of multiple phase short circuits can include widespread power outages, permanent damage to equipment, and potential safety risks.

[0172] Multiple-point ground faults: In extreme cases, multiple single-phase ground faults may occur simultaneously. This type of fault not only makes the system unstable, but may also cause the failure of protection devices, thereby increasing the severity of the fault. Multiple-point ground faults increase the difficulty of diagnosis and repair because there is more than one fault point.

[0173] Large-scale transient overvoltage: Although transient overvoltages themselves are not uncommon phenomena, large-scale transient overvoltages (such as those caused by lightning strikes or the sudden disconnection of large equipment) can cause widespread equipment damage. Such events can cause damage to substation equipment, thereby affecting the power supply of a large area.

[0174] Distributed power reverse flow: With the increase of distributed power sources (such as solar photovoltaic panels), when a grid fault occurs, the reverse current from the distributed power sources may make fault location more complicated. In this case, conventional protection measures may fail, making it difficult to isolate the fault area, thereby extending the repair time and expanding the scope of the power outage.

[0175] Failures caused by extreme weather conditions: Although failures caused by severe weather are not uncommon, damage caused by extreme weather events (such as tornadoes, hurricanes, ice storms, etc.) can be catastrophic. This type of event can cause widespread damage to infrastructure and extend the time it takes to restore power.

[0176] It should be noted that the benefit of step S300 is that the model can continuously adjust parameters during the training process to minimize the difference between the predicted value and the true value by optimizing the loss function. The reduction of the loss function means that the prediction results of the model are closer to the actual situation, thereby improving the generalization ability of the model. In the distribution network fault location task, the reduction of the loss function is directly related to the accurate identification of the fault line, which is of great significance for ensuring the stable operation of the distribution network.

[0177] S400: According to the distribution network fault model based on the improved graph neural network after training, the real-type experimental distribution line fault location is completed;

[0178] In the embodiment of the present application, once the model training is completed, it can be used for fault location. Assume that the model has been trained and the real-time operation data of the distribution network is known. When a node or line fails, the model will calculate the probability of failure of each line through the graph neural network based on the input real-time data. Assuming that the model predicts that the failure probability of line A is the highest, reaching 90%, it can be determined that line A has failed;

[0179] For example, suppose there is a distribution network, and the topology includes a power node, three load nodes (nodes 1, 2, and 3), and three lines connecting them (edges A, B, and C). The model input includes the voltage, current, and other characteristics of each node, as well as the power, current, and other characteristics of the edge. After the model training is completed, when the system detects that the voltage of a node is abnormal, the data of the node and related lines are input into the model. The results of the model output show that the failure probability of edge A is the highest (for example, 90%), so the fault can be located on edge A. At this time, the staff can take quick measures based on this information to isolate the faulty line and restore normal power supply to other lines.

[0180] In practical applications, the performance of the distribution network fault location model depends not only on the model structure and parameter settings, but also on the quality and quantity of the training data. In order to improve the generalization ability of the model, a large-scale distribution network fault simulation data set is used for training in the embodiments of the present application. These simulation data sets generate rich fault samples by introducing various fault types and scenarios into the distribution network model, covering various situations from single-point faults to multi-point faults, from simple faults to complex faults. Through the training of these simulation data sets, the model can learn the feature representations under different fault modes, thereby improving the accuracy and robustness of fault location.

[0181] In addition, in order to further improve the performance of the model, the embodiment of the present application also introduces data enhancement technology. Data enhancement generates new training samples by transforming existing training data, thereby increasing the diversity of data. For example, by slightly adjusting the topology of the distribution network, or introducing noise and interference into the simulated fault data, fault samples with different characteristics can be generated. These enhanced data help the model to be better generalized to practical applications and avoid the occurrence of overfitting.

[0182] In an optional embodiment, some edges (i.e., lines connecting two nodes) can be selected for deletion or addition based on the existing topology. Try to change the location or attributes of certain nodes, such as adjusting the location or capacity of distributed power sources. Maintain the connectivity of the adjusted topology to ensure that the adjustment does not lead to network segmentation. Random noise can also be added to measurement data such as voltage and current to simulate sensor errors. Use Gaussian distribution or other suitable distributions to add noise. Introduce random changes in known fault data to simulate the uncertainty and complexity of fault occurrence. Change the location, type or intensity of the fault to generate a variety of fault scenarios. Simulate different environmental conditions, such as temperature, humidity, etc., to observe the impact of these factors on fault detection. For example, simulate the impact of high temperature conditions on the aging rate of equipment, or the changes in equipment performance under low temperature conditions.

[0183] Specifically, assume that a load node (node ​​2) is selected for adjustment. Without changing the overall network connectivity, the connection between node 2 and another node (for example, node 3) is changed to node 1. This means that the connection between node 2 and node 3 is disconnected, and a new connection between node 2 and node 1 is created. The load of node 2 is adjusted, such as increasing or decreasing its load by a certain percentage, or changing the output power of the distributed power supply on node 2. The adjusted topology data is recorded and saved as a new training sample.

[0184] Furthermore, suppose that the edge connecting node 1 and node 2 (edge ​​A) is selected for adjustment. Gaussian noise is added to the current and voltage data of edge A. For example, assuming that the original current is 10A, Gaussian noise can be added with a standard deviation of 0.5A. Assuming that a short circuit fault occurs on edge A, additional interference can be introduced on this basis, such as changing the magnitude or duration of the short circuit current, or changing the type of short circuit (such as changing a single-phase ground short circuit to a two-phase short circuit). Record the fault data after adding noise and interference, and save it as a new training sample.

[0185] For example, assume that there is a simple distribution network with a topological structure including a power node, three load nodes (nodes 1, 2, and 3) and three lines connecting them (edges A, B, and C). The model input includes the voltage and current characteristics of each node, as well as the power and current characteristics of the edge.

[0186] The original data is:

[0187] Node 1: power node, no load, no current;

[0188] Node 2: load node, voltage 120V, current 10A;

[0189] Node 3: load node, voltage 120V, current 5A;

[0190] Edge A: connects nodes 1 and 2, current 10A;

[0191] Edge B: connects nodes 2 and 3, current 5A;

[0192] Edge C: connects node 3 and node 1, no current;

[0193] The following are examples after data augmentation:

[0194] Disconnect edge B and create a new edge D connecting nodes 1 and 3.

[0195] The load of node 2 is adjusted to 120V voltage and 12A current.

[0196] The new side D current is 12A.

[0197] When Gaussian noise (standard deviation 0.5A) is added to the current data of side A, the current becomes 10.2A.

[0198] A short circuit fault occurs on side A, and the current peak is 200A. When random fluctuations are introduced, the current peak becomes 210A.

[0199] Through the above steps, new fault samples can be generated for model training to improve the robustness and generalization ability of the model.

[0200] After the model training is completed, in order to verify the performance of the model, a strict test process is also designed in the embodiment of the present application. During the test, the fault is first simulated in the real distribution network, and then the trained model is used to locate the fault. By comparing the fault line label output by the model with the actual fault line, the positioning accuracy and response time of the model can be evaluated. In addition, in order to comprehensively evaluate the performance of the model, different fault types, different fault degrees and different environmental conditions are also considered during the test.

[0201] Finally, the distribution network fault location model in the embodiment of the present application has shown excellent performance in practical applications. The model can not only quickly and accurately locate the faulty line, but also provide the cause of the fault and possible solutions. This provides powerful decision-making support for the operation and maintenance personnel of the distribution network, and greatly improves the operating efficiency and reliability of the distribution network. Through the method in the embodiment of the present application, the distribution network fault location technology has been significantly improved, laying a solid foundation for the development of smart distribution networks.

[0202] The above is a schematic scheme of a real-type experimental distribution line fault location method of this embodiment. It should be noted that the technical scheme of the real-type experimental distribution line fault location system and the technical scheme of the above-mentioned real-type experimental distribution line fault location method belong to the same concept. The details of the technical scheme of the real-type experimental distribution line fault location system in this embodiment that are not described in detail can all be referred to the description of the technical scheme of the above-mentioned real-type experimental distribution line fault location method.

[0203] This embodiment also provides a real experimental distribution line fault location system, including:

[0204] The data acquisition module is used to obtain the topological structure diagram of the target distribution network in the real experiment;

[0205] The input acquisition module is used to obtain the node and edge features of the target distribution network as the input of the distribution network fault model to be established according to the topological structure diagram of the target distribution network in the real experiment and the preset diagram input feature selection logic;

[0206] A model training module is used to train a distribution network fault model based on an improved graph neural network according to the input of the distribution network fault model to be established and the output of a preset distribution network fault model, wherein the output of the preset distribution network fault model is the probability of line failure in the distribution network;

[0207] The fault location module is used to complete the fault location of the real experimental distribution line according to the distribution network fault model based on the improved graph neural network after training.

[0208] The above-mentioned unit modules may be embedded in or independent of a processor in a computer device in the form of hardware, or may be stored in a memory in a computer device in the form of software, so that the processor can call and execute operations corresponding to the above-mentioned modules.

[0209] The present embodiment also provides a computing device, which includes a processor, a memory, a communication interface, a display screen and an input device connected through a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The communication interface of the computer device is used to communicate with an external terminal in a wired or wireless manner, and the wireless manner can be implemented through WIFI, an operator network, NFC (near field communication) or other technologies. When the computer program is executed by the processor, a method for locating a fault in a real experimental distribution line is implemented. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen, and the input device of the computer device can be a touch layer covered on the display screen, or a button, a trackball or a touchpad provided on the housing of the computer device, or an external keyboard, touchpad or mouse, etc.

[0210] This embodiment also provides a storage medium on which a computer program is stored. When the program is executed by a processor, the method proposed in the above embodiment is implemented.

[0211] The storage medium proposed in this embodiment and the method proposed in the above embodiment belong to the same inventive concept. The technical details not fully described in this embodiment can be referred to the above embodiment, and this embodiment has the same beneficial effects as the above embodiment.

[0212] Through the above description of the implementation methods, the technicians in the relevant field can clearly understand that the present invention can be implemented by means of software and necessary general hardware, and of course can also be implemented by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as a computer floppy disk, read-only memory (ReadOnly, Memory, ROM), random access memory (RandomAccess Memory, RAM), flash memory (FLASH), hard disk or optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, server, or network device, etc.) to perform the methods of various embodiments of the present invention.

[0213] Example 2

[0214] Reference Figure 8 to Figure 13 Based on the previous embodiment, this embodiment provides an application example of a real experimental distribution line fault location method and system to verify and illustrate the technical effects used in this method.

[0215] This application has been fully analyzed and empirically studied before. In order to verify the effectiveness of the algorithm of this application, an improved single radial multi-source distribution network IEEE33 node distribution system is used for verification. The radial distribution network system contains 32 branches. In order to reflect the fault location advantage of the algorithm of this application in the multi-source distribution network, 6 distributed power sources are added to the original IEEE33 node distribution network. The topology is as follows: Figure 8 shown.

[0216] Since a large number of fault samples are required for training, this application develops a sample generation tool based on the OpenDSS (Open Distributed System Simulator) distribution network simulation software. OpenDSS is an open source three-phase distribution network power flow simulation software developed by the Electric Power Research Institute (EPRI) of the United States. OpenDSS is a software specially designed for the study of modern distribution systems, which makes up for many shortcomings of traditional distribution system analysis tools. The main function is power grid simulation, and it has a unique dynamic library (Dynamic Link Library, DLL) to meet the user's flexible modeling needs. Compared with other simulation software, OpenDSS has a richer COM (Component Object Model) interface, supports programming languages ​​such as Matlab and Python, and has a fast call response speed. Users can flexibly develop power system functions according to their needs. Fig. 9 This is the power flow diagram of the IEEE33 node established using OpenDSS. It can be seen that the line of power outflow becomes thicker at the location of the distributed power source (OpenDSS software cannot display distributed power sources):

[0217] In addition, this application builds a distribution network fault location model based on graph convolutional network based on Pytorch dynamic neural network framework. In order to better train, the OpenDSS sample generation tool is embedded in the Pytorch graph convolutional neural network program, realizing the function of generating random fault data and its labels in real time.

[0218] The distribution network topology abstraction extraction module based on OpenDSS can provide the distribution network fault location training program with an abstract distribution network topology graph structure, that is, the nodes of the distribution network and their interconnected lines are extracted as a graph; the random fault generation module first randomly selects the fault line, and then selects random points on the fault line, that is, 1% to 99% of the fault line, and selects random fault types, including 3 single-phase grounding faults, 3 two-phase faults, 3 two-phase grounding faults, and 10 types of three-phase faults; the fault node and edge data extraction module is responsible for extracting and processing the feature data on the nodes and edges, and at the same time corresponding the data to the corresponding nodes and edges in the graph. The iterative control part in the distribution network location main program mainly controls the learning process of the graph convolutional neural network; the graph convolutional neural network is used for distribution network classification; and finally the training process is recorded and output through the recording output part.

[0219] The graph convolutional neural network of this application is divided into an input layer, a hidden layer, and an output layer. The input layer aggregates and splices the adjacent edge features of each node to form a merged node feature. The number of neurons in the input layer is 6, that is, the output dimension of the merged node feature is 6; the hidden layer adopts the GraphSAGE algorithm based on the LSTM aggregator, with a total of three hidden layers of GraphSAGE neural network, and the output node feature dimensions of each layer from bottom to top are 10, 20, and 10 respectively; the output layer aggregates the node features to the edge, and each edge establishes a fully connected layer to output the 20-dimensional features of the nodes on both sides to the edge features of 3 dimensions, and finally outputs them through the fully connected layers of all edges and the softmax activation function, and the output is a 32-dimensional classification result. The learning adopts the cross entropy error loss function and the Adam optimizer.

[0220] The experiment selected different input characteristics for testing. Considering that it may be difficult to accurately measure the phase angle synchronously in the actual distribution network, this application takes Fig.10 The three methods in are tested as input.

[0221] The sequence voltage amplitude is the positive, negative and zero sequence voltage amplitude, and the same is true for the sequence current. The edge overcurrent alarm information is mainly the overcurrent information fed back by the FTU device on each line. The forward overcurrent line is 1, the reverse overcurrent line is -1, and no overcurrent is 0.

[0222] The average accuracy in the figure is the average accuracy of nearly 100 positionings, among which the node sequence voltage amplitude and sequence current amplitude inputs reached an accuracy of 95.1% after 200,000 sample training times. Although it is basically close to convergence, it still has a steady increase. It reached an accuracy of 97.4% at 430,000 times not shown in the figure. For the convenience of observation, this application intercepted the data at 200,000 times. It can be seen from the figure that the average positioning accuracy of the sequence voltage and current is higher than that of other input quantities.

[0223] according to Fig.11 As shown in the figure, a fully connected neural network, a GraphSAGE spatial domain graph convolution network, and a GCN spectral domain graph convolution network with the same number of layers are constructed, and the number of neurons in each layer of the fully connected neural network is set to be greater than the sum of the dimensions of all nodes in the graph convolution to ensure that the parameters of the fully connected neural network are not less than those of the graph convolution network. The same samples are used to train them respectively, and it can be seen that the GraphSAGE algorithm is about 10% more accurate than the GCN algorithm, and the accuracy of both algorithms is much higher than that of the fully connected neural network.

[0224] The aggregation function comparison of the GraphSAGE algorithm is as follows Fig.12 As shown ( Fig.10 , Fig.11 The left vertical coordinate in Fig.12 The pool aggregation function has a higher accuracy than the LSTM aggregation function, reaching 97.9% accuracy at 200,000 training times. The LSTM aggregation function is slightly higher than the mean aggregation function overall, and the result is more stable than the mean aggregation function.

[0225] Fault transition resistance has a great influence on the fault location of the distribution network. This application uses two random faults in different intervals to train the model, where model A is trained with a random transition resistance value of 0.2 to 15Ω, and model B is trained with a random transition resistance value of 0.2 to 3000Ω. A total of 100,000 data are extracted from 330,600 to 430,600 trainings of the two models for statistics. When the two models converged at 430,600 times, the average accuracy of the thousand data was 98.18% and 97.60% respectively. The positioning accuracy of transition resistances of different sizes is shown in Table 1 below:

[0226] Table 1: Comparison of positioning accuracy of transition resistors of different sizes.

[0227]

[0228] It can be concluded that the model has the highest accuracy at the intermediate transition resistance. The positioning accuracy of model A for small transition resistance is higher than that of model B, but it is not as good as model B for large transition resistance.

[0229] The algorithm of this application is applied to the IEEE 33-node distribution network with the immune algorithm and the traditional neural network algorithm. The average accuracy of the model A of this application during 430,600 trainings is compared. The results are shown in Table 2:

[0230] Table 2: Comparison of accuracy with traditional algorithms.

[0231]

[0232] Traditional immune algorithms use specific fault samples for iterative calculations, and the samples do not contain distributed power supplies. The algorithm in this application uses randomly generated fault samples of various types for calculations. The accuracy of the algorithm in this application is 98.18% based on 2 hours and 430,600 training times. Although the accuracy is basically stable, it is still rising slowly. Due to the relationship between time and data volume, this application did not continue training, and the accuracy of this application still has room for improvement.

[0233] Fig.13 is the loss function of various algorithms. It can be seen that GraphSAGE and GCN have good training convergence ability in the distribution network fault location problem of this application. The above experimental results show that the graph convolution network has a strong advantage in processing distribution network graph structure data, and can locate the distribution network fault position more accurately. The training sample data of this application has extremely strong randomness. The results prove that the distribution network fault location algorithm based on graph convolution network has high accuracy and robustness, and the training process stability and convergence ability are good.

[0234] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

[0235] Those skilled in the art will appreciate that the embodiments of the present application can be provided as methods, systems, or computer program products. Therefore, the present application can adopt the form of complete hardware embodiments, complete software embodiments, or embodiments in combination with software and hardware. Moreover, the present application can adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that contain computer-usable program code. The scheme in the embodiments of the present application can be implemented in various computer languages, for example, object-oriented programming language Java and literal scripting language JavaScript, etc.

[0236] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0237] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0238] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process in the computer or other programmable device. Figure 1 A process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0239] Although the preferred embodiments of the present application have been described, those skilled in the art may make other changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications falling within the scope of the present application.

[0240] Obviously, those skilled in the art can make various changes and modifications to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the claims of the present application and their equivalents, the present application is also intended to include these modifications and variations.

Claims

1. A real-type experimental distribution line fault location method, characterized in that: include: Obtain the topological structure diagram of the target distribution network in the real experiment; According to the topological structure diagram of the target distribution network in the real experiment, combined with the preset graph input feature selection logic, the node and edge features of the target distribution network are obtained as the input of the distribution network fault model to be established; According to the input of the distribution network fault model to be established and the output of the preset distribution network fault model, training the distribution network fault model based on the improved graph neural network, wherein the output of the preset distribution network fault model is the probability of line failure in the distribution network; According to the distribution network fault model based on the improved graph neural network after training, the real experimental distribution line fault location is completed.

2. The real-type experimental distribution line fault location method according to claim 1, characterized in that: The preset graph input feature selection logic includes: The target distribution network node characteristics include six-dimensional three-phase voltage amplitude, phase angle positive sequence, negative sequence, zero sequence voltage amplitude, phase angle, two-dimensional neutral point voltage amplitude, phase angle, and six-dimensional three-phase injection / output active and reactive power; The target distribution network edge features include six-dimensional three-phase line current amplitude, phase angle, positive sequence, negative sequence, zero sequence current amplitude, phase angle, six-dimensional three-phase line power amplitude, phase angle, and one-dimensional fault current alarm information; Determine the model diagram input features according to the automation degree of the target distribution network, the type of the measuring device and the data category of the information source of the target distribution network; Among them, all nodes are required to have the same input dimension, and all edges are required to have the same input dimension.

3. The real-type experimental distribution line fault location method according to claim 2, characterized in that: The distribution network fault model based on the improved graph neural network includes: an input layer, a hidden layer, and an output layer; The input layer receives the model graph input features and outputs the processed graph input features to the hidden layer. The graph convolution algorithm used by the hidden layer is based on node feature data transmission and does not accept any edge feature data. If the model graph input feature only selects the data of the target distribution network edge feature, the data of the target distribution network edge feature is directly input into the hidden layer; If the model graph input feature contains data of the edge feature of the target distribution network, it must be formatted, the edge feature data must be converted into node feature data, and then output to the hidden layer; The input layer includes merging the original edge features into node features through a single-layer neural network.

4. The real-type experimental distribution line fault location method according to claim 3, characterized in that: The input layer includes merging the original edge features into node features through a single-layer neural network, including: According to the dimension d of the edge feature vector e , construct a single-layer fully connected neural network and add a bias; Input edge feature data E1 and E2 of each edge into the single-layer fully connected neural network to obtain output values ​​E′1 and E′2 of the same dimension; Sum the neural network output vectors of all edges to get the sum vector E of the same dimension SUM ; The node feature vector N V and the edge sum vector E SUM By Concat(N V , E SUM ) operations are merged into a new vector as output.

5. The real-type experimental distribution line fault location method according to claim 3, characterized in that: The hidden layer includes: Building a multi-layer graph convolutional neural network in the hidden layer; The multi-layer graph convolutional neural network is constructed by stacking several single hidden layers, with node v i For example, the node features of the single hidden layer in the l+1 layer The aggregation formula is: Among them, N(v i ) is the set of neighbor nodes, is the node v in the lth layer j The node feature vector of For node v i The result vector after the neighbors are aggregated, aggregate l+1 (·) is the learnable aggregation function of the l+1th layer, concat(·) is the concatenation function, which concatenates two vectors horizontally in their original dimensions, σ(·) is the Relu activation function, and W is the learnable weight parameter matrix.

6. The real-type experimental distribution line fault location method according to claim 3, characterized in that: The output layer includes: The node feature data in the hidden layer are merged, and the edge features are output through the fully connected layer: AND 1,2 =σ(W 1,2 ·Concat(V1,V2)) Among them, V1 and V2 are the node feature vectors of the two end points of the edge (v1, v2), which are passed in from the hidden layer. 1,2 is the feature vector of the processed output edge (v1, v2), Concat(V1, V2) is the merging of the hidden layer feature vectors of nodes v1 and v2, W 1,2 is an m×n parameter matrix, σ(·) is the ReLU activation function, which outputs a feature vector with a length of m for each edge, and n represents the number of edges; The number of outputs of the fully connected layer is the number of classifications of the distribution network fault identification task, that is, the number of lines in the distribution network. Each value in the output vector represents the probability of the corresponding line failure, and the sum of all values ​​is 1.

7. The real-type experimental distribution line fault location method according to claim 2, characterized in that: The distribution network fault model based on the improved graph neural network also includes: The loss function calculation formula in the distribution network fault model based on the improved graph neural network is: Among them, T is the label vector, Y is the predicted output vector, and y a is the probability of the ath fault predicted by the model, w a The weight factor corresponding to the ath fault type is W1=[w1,w2,...,w n ] is the weight vector, A represents the total number of faults, t a represents the true label.

8. A system using the real-type experimental distribution line fault location method as claimed in any one of claims 1 to 7, characterized in that: include: The data acquisition module is used to obtain the topological structure diagram of the target distribution network in the real experiment; The input acquisition module is used to obtain the node and edge features of the target distribution network as the input of the distribution network fault model to be established according to the topological structure diagram of the target distribution network in the real experiment and the preset diagram input feature selection logic; A model training module, used to train a distribution network fault model based on an improved graph neural network according to the input of the distribution network fault model to be established and the output of a preset distribution network fault model, wherein the output of the preset distribution network fault model is the probability of line failure in the distribution network; The fault location module is used to complete the fault location of the real experimental distribution line according to the distribution network fault model based on the improved graph neural network after training.

9. An electronic device, comprising: Memory and processor; The memory is used to store computer executable instructions, and the processor is used to execute the computer executable instructions. When the computer executable instructions are executed by the processor, the steps of the method according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having computer program instructions stored thereon, wherein the computer-executable instructions, when executed by a processor, implement the steps of the method according to any one of claims 1 to 7.

Images