Power distribution network fault line selection method considering topological structure change

Abstract

The invention relates to the technical field of electric power operation and maintenance, and particularly discloses a power distribution network fault line selection method considering topological structure change, and the method comprises the steps: collecting zero-sequence current data, converting the zero-sequence current data into graph structure data, and providing input information representing the topology and current characteristics of a power distribution network for a model; learning the adjacency matrix and the node features by using a graph attention network until the model can accurately predict the probability of the single-phase earth fault of the power distribution network; based on the node fault probability output by the model, a threshold value is set to identify and position a specific fault line; and inputting field zero-sequence current data into the trained model so as to quickly and accurately judge the fault position of the power distribution network. According to the power distribution network fault line selection method considering topological structure changes, the efficiency and accuracy of fault line selection can be improved, and therefore effective fault monitoring can be achieved in a wider power distribution network environment.

Description

Technical Field

[0001] The present invention relates to the technical field of electric power operation and maintenance, and in particular to a distribution network fault line selection method considering topology changes. Background Art

[0002] The safe and reliable operation of the distribution network is an important foundation for realizing the national strategy of "developing smart grids and distributed energy". There has long been a problem of distribution network fault detection (line selection, positioning) at home and abroad. Fault detection (line selection, positioning) of the distribution network is one of the keys to ensure the safety of power distribution. However, since the distribution network usually contains multiple branches and nodes according to power demand and has a complex topological structure, high-resistance grounding faults such as tree barriers, incomplete breakdown of lightning arresters, and conductor drops occur frequently on distribution lines. If the system operates with a high-resistance grounding fault for a long time, it may also cause a two-phase grounding short circuit fault, expand the scope and nature of the fault, especially the fall of the conductor to the ground will cause a greater personal safety hazard. Therefore, for high-resistance grounding faults, it is also necessary to accurately detect (line selection, positioning) and perform corresponding treatments such as arc extinguishing, overvoltage protection, and even cutting off the fault line or section.

[0003] At present, the access of distributed power sources and new lines makes the topology of the distribution network more complex and unstable. The existing fault line selection and location methods are no longer suitable for complex network environments containing these factors. Among them, although the fault line selection method based on traveling waves has a fast response, it is easily affected by the terrain and line structure, resulting in a decrease in the accuracy of line selection; the fault line selection method based on the action information of the protection device depends on the setting of the threshold of the protection device, and often causes false refusal, false alarm, etc.; the fault line selection method based on artificial intelligence often relies on a fixed network structure and pre-set parameters. This dynamic change brings additional difficulty to fault detection, because the fault characteristics may change due to changes in the network structure, which reduces the accuracy and reliability of such methods. In addition, although most existing fault detection methods can accurately determine the location of the fault, they lack a comprehensive consideration of the topological structure of the entire distribution network. These methods cannot make full use of the structural information of the distribution network to optimize the fault line selection process, thereby affecting the comprehensiveness and depth of fault line selection.

[0004] In order to meet these challenges, the present invention is committed to developing a distribution network fault line selection method that takes into account changes in topology structure. It can not only quickly and accurately locate the fault line, but also comprehensively consider the topological structure information of the distribution network and adaptively adjust to match the dynamically changing structure of the distribution network, thereby improving the efficiency and accuracy of fault line selection, thereby achieving effective fault monitoring in a wider distribution network environment. Summary of the invention

[0005] One object of the present invention is to provide a distribution network fault line selection method that takes into account changes in topology structure, which can not only quickly and accurately locate the fault line, but also comprehensively consider the topological structure information of the distribution network and adaptively adjust to match the dynamically changing structure of the distribution network, thereby improving the efficiency and accuracy of fault line selection, thereby achieving effective fault monitoring in a wider distribution network environment.

[0006] To achieve the above objectives, the present invention provides a distribution network fault line selection method considering topology changes, comprising:

[0007] S10: Data preprocessing: by collecting zero-sequence current data and converting it into graph structure data, the model is provided with input information characterizing the distribution network topology and current characteristics;

[0008] S20: Model training: Use the graph attention network to learn the adjacency matrix and node features until the model can accurately predict the probability of single-phase grounding faults in the distribution network;

[0009] S30: Fault judgment: Based on the node failure probability output by the model, a threshold is set to identify and locate the specific fault line;

[0010] S40: Model application: Input the on-site zero-sequence current data into the trained model to quickly and accurately determine the fault location of the distribution network.

[0011] Optionally, the S10 includes:

[0012] S101: Data collection: collecting zero-sequence current data from the distribution network site;

[0013] S102: Data cleaning: removing noise and abnormal values ​​from the collected zero-sequence current data;

[0014] S103: Feature extraction: extract key features from the cleaned data;

[0015] S104: Constructing a graph structure: converting the node and branch information of the distribution network into graph structure data, where the nodes represent various parts of the distribution network and the edges represent the connection relationship between the nodes.

[0016] Optionally, the key features extracted from the data include current amplitude and / or phase angle.

[0017] Optionally, the S104 includes:

[0018] First, the topological connection relationship of the distribution network measurement nodes is obtained, and a node connection matrix is ​​made according to formula (1). The zero-sequence current data samples required for the experiment are obtained by measuring the data of the zero-sequence current transformer in the field.

[0019] The sample data obtained from field measurements are normalized using formula (3) to normalize the fault current signal of each measuring point, and the processed experimental data are obtained, which are used as the node characteristic data of each node. The nodes on both sides of the fault section of the IEEE18-node network are regarded as fault nodes, and the label of the fault node is set to 1, and the label of the non-fault node is set to 0:

[0020] Specifically:

[0021] From the perspective of graph, without considering the internal components and line characteristics, the distribution network node network is abstracted as a graph structure data G = (X, A) of nodes and edges; the adjacency matrix A describes the connection relationship between the nodes of the distribution network, where the row and column numbers of the adjacency matrix correspond to the nodes, as shown in formula (1):

[0022] (1);

[0023] The branch zero-sequence current is used as the characteristic vector of the branch head node; in the power system, the zero-sequence current refers to the current generated when the vector sum of the three-phase currents is not zero in a three-phase system; the zero-sequence current is usually related to ground faults; in actual scenarios, for cable lines, the zero-sequence current transformer is a device specifically used to detect zero-sequence current. It is installed at the neutral point of the three-phase power system and can detect the zero-sequence current generated by faults or unbalanced loads; the zero-sequence current transformer is used to collect zero-sequence current; and the current measurement of overhead lines usually uses three-phase current transformers, which are installed on each line of the three phases respectively; by measuring the three-phase current, the zero-sequence current can be calculated; the calculation method is as follows:

[0024] (2);

[0025] in, , and They are the three-phase currents A, B and C measured by the three-phase current transformer;

[0026] X is the node feature vector: (3);

[0027] is the number of nodes; Contains this node The sampling point time series samples are 1024 in this embodiment; the normalization method is used to preprocess the data, and the formula is as follows:

[0028] (4);

[0029] in, is the normalized data, is the minimum value in the original data, is the maximum value in the original data; after this processing, the minimum value in the original data is mapped to 0, the maximum value is mapped to 1, and other values ​​are linearly mapped between 0 and 1.

[0030] Optionally, the S20 is specifically:

[0031] The processed adjacency matrix and node features are input into the constructed graph neural network model. The spatial features and temporal features are extracted respectively through GAT and LSTM. Different weights are assigned to different nodes through the graph attention mechanism to aggregate more obvious spatial features. At the same time, the multi-gate structure unit of LSTM captures multiple dependencies of timing signals through forgetting, updating and outputting. The time and space feature vectors are spliced ​​together through a splicing layer. After passing through a fully connected layer and a sigmoid classification layer, the prediction results and the actual fault labels are calculated using the cross entropy loss function. According to the size of the loss value, the model parameters are updated through back propagation to optimize the model in the direction of smaller loss value. According to the probability Pi of each node failure output by the model, it is determined between which two nodes of the distribution network there is a fault, thereby determining the fault section. When the training result meets a certain accuracy, the model parameters are saved to obtain a trained model.

[0032] Optionally, the S20 includes:

[0033] S201: Initialize the model: set the initial parameters of the graph attention network model, including weights and biases;

[0034] S202: Input data: The adjacency matrix and node features of the constructed graph structure data are used as input data;

[0035] S203: Forward propagation: Calculate the correlation between nodes through the graph attention mechanism and update the node features;

[0036] S204: Loss function: Define a loss function to measure the difference between the model output and the actual fault location;

[0037] S205: Back propagation: Calculate the gradient according to the loss function and update the model parameters;

[0038] S206: Iterative optimization: repeat the forward propagation and back propagation process until the model loss reaches a preset threshold or the number of training times reaches a set value;

[0039] S207: Model saving: saving the trained model parameters for subsequent application.

[0040] Optionally, the S203 includes:

[0041] The graph attention network structure is divided into two main parts, namely GAT and LSTM. Among them, GAT extracts spatial features and LSTM extracts temporal features. GAT introduces an attention mechanism based on GCN to set weights for the information aggregation of nodes and their adjacent nodes, thereby aggregating more significant features of each node. The input of each layer of GAT is a set of feature vectors of the node. ; N is the number of topological nodes in the distribution network; x i is the feature vector of node i; F is the feature number of the node; its output is a new feature set: Where x and Has different dimensions;

[0042] First, calculate the attention coefficient between each node, expressed as

[0043] (5);

[0044] where e ij is the attention coefficient of node j (neighboring node of node i) to node i; is the attention function, which maps the concatenated high-dimensional features to a real number; •||• is the feature concatenation operation; x i and x j is the input feature vector of node i and node j; W is the shared parameter matrix of linear transformation; N i is the total number of first-order neighbor nodes of node i; the calculated attention coefficient is nonlinearly transformed and normalized, as shown in the following formula

[0045] (6);

[0046] After that, the adjacent node vectors are weighted and summed according to the attention weights to update the feature vector of the target node i, as shown in the following formula:

[0047] (7);

[0048] Use a multi-head attention mechanism to improve the performance of the model, and set the number of attention heads to 4; that is, use 4 independent attention mechanisms to calculate the attention weights and update the node features, and then concatenate the vectors updated by each attention mechanism;

[0049] LSTM is a variant of recurrent neural network, which makes up for the shortcomings of recurrent neural network that it cannot remember information from a long time ago and the gradient will explode and disappear. The LSTM hidden layer is formed by stacking multiple gate structures. Each gate structure contains an input gate, an output gate and a forget gate, as well as two functions sigmod and tanh.

[0050] The calculation formula for each time step is as follows:

[0051] (8);

[0052] In the formula is the output of the forget gate, , are the weights corresponding to the output of the previous gate structure and the input of the gate structure in the current period, is the output of the gate structure in the previous period, is the input of the gate structure of the current period, is the bias term; (sigmod) The value of each dimension output is between 0 and 1, 1 means that the original information is completely retained, and 0 means that the original information is completely removed;

[0053] (9);

[0054] (10);

[0055] In formula (9) and (10) is the activation vector of the input gate, is the information accumulated at the current moment; multiply the two together to get the state quantity to be updated;

[0056] (11);

[0057] In formula (11), The state input of the next gate structure is obtained by The state information that needs to be forgotten with the current structure Multiply and add the state amount that needs to be updated in the current state The state quantity of the next gate structure input can be obtained;

[0058] (12);

[0059] (13);

[0060] Formula (12) is the output gate activation vector of LSTM, and formula (13) is the output of LSTM. It multiplies the activation vector of the output gate and the state of the current gate structure by tanh to get the output.

[0061] Optionally, the S30 includes:

[0062] S301: Fault probability calculation: Use the trained model to predict the fault probability of the new graph structure data;

[0063] S302: Fault location: According to the output node fault probability Pi, determine the section where the fault is most likely to occur;

[0064] S303: Threshold setting: a fault probability threshold is set. Only when the fault probability of a node exceeds the threshold, the node is considered to be faulty.

[0065] Optionally, the S204 includes:

[0066] The graph attention network aggregates the input adjacency matrix and node features through GAT for spatial feature aggregation, and uses LSTM in parallel with GAT to extract the time series features of zero-sequence current of each node; the feature vectors obtained by the two are spliced ​​so that each node has both the spatial features of the distribution network topology and the time series features of zero-sequence current. The features are nonlinearly transformed through two fully connected layers, and finally the probability p of each node being a faulty node is output through the Sigmoid activation function. i , thereby determining the fault section; the Sigmoid function formula is as follows:

[0067] (14);

[0068] in is the input value, e is the base of natural logarithm;

[0069] The output prediction result is compared with the original label, and the loss value is calculated using the Focal Loss loss function. FocalLoss adds an adjustment factor to reduce the weight of easy-to-classify samples on the basis of the balanced cross entropy loss function, focusing on the training of difficult samples. Its definition is as follows:

[0070] (15);

[0071] in, is the probability value predicted by the model, is the balancing factor, is a regulating factor; and The value of can make Focal Loss pay more attention to difficult samples, thereby improving the performance of the model on unbalanced data sets.

[0072] Optionally, the S40 includes:

[0073] S401: Data input: After preprocessing, the zero-sequence current data collected on site is input into the trained model;

[0074] S402: Fault prediction: The model outputs the fault probability of each node and sorts them according to the probability;

[0075] S403: Fault line determination: Determine the fault line in the distribution network according to the node with the highest fault probability;

[0076] S404: Result output: output the fault line information.

[0077] The beneficial effects of the present invention are: providing a distribution network fault line selection method that takes into account changes in topological structure, abstracting the topological structure of the distribution network into a graph structure composed of nodes and edges, and using the zero-sequence current of the distribution network as the feature vector of each node. The graph structure information is converted into an adjacency matrix, and used as a model input together with the zero-sequence current of each node. The graph attention mechanism of the GAT network is used to aggregate the spatial features of each node, and the LSTM network extracts the time features of each node, and the spatial features and the time features are spliced ​​to form new node features. The obtained node features are used to distinguish between faulty nodes and non-faulty nodes, thereby achieving the effect of locating the faulty line. Through the graph attention mechanism of GAT, different weights are assigned to different neighborhood nodes to achieve fault line selection when the topological structure of the distribution network changes.

[0078] Therefore, the distribution network fault line selection method considering topology changes provided by the present invention can not only quickly and accurately locate the fault line, but also comprehensively consider the topology information of the distribution network and adaptively adjust to match the dynamically changing structure of the distribution network, thereby improving the efficiency and accuracy of fault line selection, thereby achieving effective fault monitoring in a wider distribution network environment. BRIEF DESCRIPTION OF THE DRAWINGS

[0079] In order to more clearly illustrate the technical solutions in this embodiment or the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0080] Figure 1 It is a framework diagram of a graph attention network model provided by an embodiment of the present invention;

[0081] Figure 2 is a topological structure diagram of an IEEE18 distribution network provided in an embodiment of the present invention;

[0082] Figure 3 A flow chart of a distribution network fault line selection method considering topology changes provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0083] In order to make the purpose, features and advantages of the present invention more obvious and easy to understand, the technical scheme in this embodiment will be clearly and completely described below in conjunction with the drawings in this embodiment. Obviously, the embodiments described below are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0084] This embodiment provides a distribution network fault line selection method that takes into account changes in topology structure, which is suitable for single-phase grounding fault detection and line selection scenarios under changes in distribution network topology structure, and can improve the efficiency of power grid operation monitoring. The distribution network fault line selection method that takes into account changes in topology structure is executed by a fault line selection device and is implemented through software and / or hardware.

[0085] Figure 3 This is a flow chart of the distribution network fault line selection method considering topology changes provided in the first embodiment.

[0086] See also Figure 3 The distribution network fault line selection method considering topology changes comprises the following steps:

[0087] S10: Data preprocessing: by collecting zero-sequence current data and converting it into graph structure data, the model is provided with input information characterizing the distribution network topology and current characteristics;

[0088] S101: Data collection: collecting zero-sequence current data from the distribution network site;

[0089] S102: Data cleaning: removing noise and abnormal values ​​from the collected zero-sequence current data;

[0090] S103: Feature extraction: extract key features from the cleaned data, such as current amplitude, phase angle, etc.;

[0091] S104: constructing a graph structure: converting the node and branch information of the distribution network into graph structure data, where the nodes represent various parts of the distribution network and the edges represent the connection relationship between the nodes;

[0092] First, the topological connection relationship of the distribution network measurement nodes is obtained, and a node connection matrix is ​​made according to formula (1). The zero-sequence current data samples required for the experiment are obtained by measuring the data of the zero-sequence current transformer in the field.

[0093] From the perspective of graph, without considering the internal components and line characteristics, the distribution network node network is abstracted as a graph structure data G = (X, A) of nodes and edges; the adjacency matrix A describes the connection relationship between the nodes of the distribution network, where the row and column numbers of the adjacency matrix correspond to the nodes, as shown in formula (1):

[0094] (1);

[0095] The branch zero-sequence current is used as the characteristic vector of the branch head node. In the power system, zero-sequence current refers to the current generated when the vector sum of the three-phase currents is not zero in a three-phase system. Zero-sequence current is usually associated with ground faults. In actual scenarios, for cable lines, zero-sequence current transformers are devices specifically used to detect zero-sequence current. They are usually installed at the neutral point of the three-phase power system and can detect zero-sequence currents caused by faults or unbalanced loads. Zero-sequence current transformers are generally used to collect zero-sequence current. The current measurement of overhead lines usually uses three-phase current transformers, which are installed on each of the three-phase lines. By measuring the three-phase current, the zero-sequence current can be calculated. The calculation method is as follows:

[0096] (2);

[0097] in, , and They are the A, B, and C three-phase currents measured by the three-phase current transformer.

[0098] X is the node feature vector: (3);

[0099] is the number of nodes; Contains this node The sampling point time series samples are 1024 in this embodiment. The normalization method is used to preprocess the data, and the formula is as follows:

[0100] (4);

[0101] in, is the normalized data, is the minimum value in the original data, is the maximum value in the original data. After this processing, the minimum value in the original data is mapped to 0, the maximum value is mapped to 1, and other values ​​are linearly mapped between 0 and 1. This helps to reduce the impact of outliers on the model.

[0102] Optionally, the sample data obtained by field measurement is used to normalize the fault current signal of each measuring point using formula (3) to obtain processed experimental data, which is used as the node characteristic data of each node. Figure 2 The nodes on both sides of the faulty segment are treated as faulty nodes, and the labels of the faulty nodes are set to 1, and the labels of the non-faulty nodes are set to 0.

[0103] S20: Model training: Use the graph attention network to learn the adjacency matrix and node features until the model can accurately predict the probability of single-phase grounding faults in the distribution network;

[0104] S20 is specifically as follows: the processed adjacency matrix and node features are input into the constructed graph neural network model, spatial features and temporal features are extracted respectively through GAT and LSTM, different weights are assigned to different nodes through the graph attention mechanism to aggregate more obvious spatial features, and at the same time, the multi-gate structure unit of LSTM captures multiple dependencies of timing signals through forgetting, updating and outputting, and splices the temporal and spatial feature vectors together through a splicing layer. After passing through a fully connected layer and a sigmoid classification layer, the obtained prediction results and the actual fault labels are calculated using the cross entropy loss function. According to the size of the loss value, the model parameters are updated through back propagation to optimize the model in the direction of smaller loss value; according to the probability P of each node failure output by the model i , determine between which two nodes of the distribution network there is a fault, and thus determine the fault section; when the training result meets a certain accuracy, save the model parameters and obtain the trained model;

[0105] Specifically:

[0106] S201: Initialize the model: set the initial parameters of the graph attention network model, including weights and biases;

[0107] Using graph theory knowledge, the zero-sequence current signal characteristics of each node in the distribution network are extracted and integrated with the distribution network topology structure, and a distribution network ground fault line selection model based on graph attention network is established. The input of the graph attention network is divided into two parts, one is the adjacency matrix representing the topology of the distribution network, and the other is the node characteristics corresponding to each node, namely the zero-sequence current.

[0108] S202: Input data: The adjacency matrix and node features of the constructed graph structure data are used as input data;

[0109] S203: Forward propagation: Calculate the correlation between nodes through the graph attention mechanism and update the node features;

[0110] See also Figure 1 , the graph attention network structure is divided into two main parts, namely GAT and LSTM. Among them, GAT extracts spatial features and LSTM extracts temporal features. Among them, GAT introduces the attention mechanism based on GCN to set weights for the information aggregation of nodes and their adjacent nodes, thereby aggregating more significant features of each node. The input of each layer of GAT is a set of feature vectors of the node . N is the number of topological nodes in the distribution network; x iis the feature vector of node i; F is the feature number of the node. Its output is a new feature set: Where x and Has different dimensions.

[0111] First, calculate the attention coefficient between each node, expressed as

[0112] (5);

[0113] where e ij is the attention coefficient of node j (neighboring node of node i) to node i; is the attention function, which maps the concatenated high-dimensional features to a real number; •||• is the feature concatenation operation; x i and x j is the input feature vector of node i and node j; W is the shared parameter matrix of linear transformation; N i is the total number of first-order neighbor nodes of node i. The calculated attention coefficient is nonlinearly transformed and normalized as follows:

[0114] (6);

[0115] After that, the adjacent node vectors are weighted and summed according to the attention weights to update the feature vector of the target node i, as shown in the following formula:

[0116] (7);

[0117] A multi-head attention mechanism is used to improve the performance of the model, and the number of attention heads is set to 4. That is, the attention weights are calculated and the node features are updated through 4 independent attention mechanisms, and then the vectors updated by each attention mechanism are concatenated.

[0118] When the topology of the distribution network changes, such as the addition or removal of nodes or branches, this will cause the network's adjacency matrix and node feature vectors to change accordingly. In this case, the graph attention network (GAT) can automatically adjust the attention weights between nodes to adapt to these changes through its attention mechanism, as shown in formula (4). Since the operation of GAT is performed independently for each vertex, its core parameter W and mapping function α mainly depend on the characteristics of the node, which means that GAT can flexibly handle changes in the topology and maintain the applicability and effectiveness of the model. In short, the design of GAT allows it to dynamically adapt to changes in the network structure without retraining or complex adjustments.

[0119] LSTM is a variant of recurrent neural network, which makes up for the shortcomings of recurrent neural network that it cannot remember information from a long time ago and the gradient will explode and disappear. The LSTM hidden layer is formed by stacking multiple gate structures, each of which contains an input gate, an output gate and a forget gate, as well as two functions sigmod and tanh.

[0120] The calculation formula for each time step is as follows:

[0121] (8);

[0122] In the formula is the output of the forget gate, , are the weights corresponding to the output of the previous gate structure and the input of the gate structure in the current period, is the output of the gate structure in the previous period, is the input of the gate structure of the current period, is the bias term; The value of each dimension of the (sigmod) output is between 0 and 1, 1 means that the original information is completely retained, and 0 means that the original information is completely removed.

[0123] (9);

[0124] (10);

[0125] In formula (9) and (10) is the activation vector of the input gate, is the information accumulated at the current moment. Multiplying the two together gives the state quantity to be updated.

[0126] (11);

[0127] In formula (11), The state input of the next gate structure is obtained by The state information that needs to be forgotten with the current structure Multiply and add the state amount that needs to be updated in the current state The state quantity of the next gate structure input can be obtained.

[0128] (12);

[0129] (13);

[0130] Formula (12) is the output gate activation vector of LSTM, and formula (13) is the output of LSTM. It multiplies the activation vector of the output gate and the state of the current gate structure by tanh to get the output.

[0131] S204: Loss function: Define a loss function to measure the difference between the model output and the actual fault location;

[0132] The graph attention network aggregates the input adjacency matrix and node features through GAT for spatial feature aggregation, and uses LSTM in parallel with GAT to extract the time series features of zero-sequence current of each node. The feature vectors obtained by the two are concatenated so that each node has both the spatial features of the distribution network topology and the time series features of zero-sequence current. The features are transformed nonlinearly through two fully connected layers, and finally the probability p of each node being a faulty node is output through the Sigmoid activation function. i , thus determining the fault section. The Sigmoid function formula is as follows:

[0133] (14);

[0134] in is the input value and e is the base of natural logarithms.

[0135] The output prediction result is compared with the original label, and the loss value is calculated using the Focal Loss loss function. FocalLoss adds an adjustment factor to reduce the weight of easy-to-classify samples on the basis of the balanced cross entropy loss function, focusing on the training of difficult samples. Its definition is as follows:

[0136] (15);

[0137] in, is the probability value predicted by the model, is the balancing factor, is the adjustment factor. and The value of can make Focal Loss pay more attention to difficult samples, thereby improving the performance of the model on unbalanced data sets.

[0138] S205: Back propagation: Calculate the gradient according to the loss function and update the model parameters;

[0139] Back propagation updates the gradient, making the model performance approach the direction of smaller loss value.

[0140] S206: Iterative optimization: Repeat the forward propagation and back propagation processes until the model loss reaches a preset threshold or the number of training times reaches a set value.

[0141] S207: Model saving: saving the trained model parameters for subsequent application.

[0142] S30: Fault judgment: Based on the node failure probability output by the model, a threshold is set to identify and locate the specific fault line;

[0143] S301: Fault probability calculation: Use the trained model to predict the fault probability of the new graph structure data;

[0144] S302: Fault location: According to the output node fault probability Pi, determine the section where the fault is most likely to occur;

[0145] S303: Threshold setting: a fault probability threshold is set. Only when the fault probability of a node exceeds the threshold, the node is considered to be faulty.

[0146] S40: Model application: Input the on-site zero-sequence current data into the trained model to quickly and accurately determine the fault location of the distribution network;

[0147] S401: Data input: After preprocessing, the zero-sequence current data collected on site is input into the trained model;

[0148] S402: Fault prediction: The model outputs the fault probability of each node and sorts them according to the probability;

[0149] S403: Fault line determination: Determine the fault line in the distribution network according to the node with the highest fault probability;

[0150] S404: Result output: The fault line information is output for reference and processing by the operation and maintenance personnel.

[0151] The method for selecting a fault line in a distribution network considering changes in topology structure provided in this embodiment has the following beneficial effects:

[0152] ① Improve fault detection accuracy: By combining the graph attention network (GAT) and the long short-term memory network (LSTM), this method can simultaneously capture the spatial characteristics and temporal characteristics of the distribution network, thereby more accurately identifying and locating single-phase grounding faults.

[0153] ② Adapt to changes in topology: This method can adapt to changes in the topology of the distribution network, such as the addition or removal of nodes or branches, without the need to retrain the model or make complex adjustments, thereby improving the flexibility and adaptability of the model.

[0154] ③ Reduce operation and maintenance costs: By quickly and accurately determining the fault location, the workload of operation and maintenance personnel can be reduced, operation and maintenance costs can be reduced, and the efficiency of power grid operation monitoring can be improved.

[0155] ④ Enhance grid reliability: Timely and accurate detection and location of faults can quickly take measures to restore normal grid operation, reduce power outage time, and improve grid reliability and user satisfaction.

[0156] ⑤ Handling unbalanced data sets: By using the Focal Loss loss function, this method can better handle unbalanced data sets, focus on the training of difficult samples, and improve the performance of the model on unbalanced data sets.

[0157] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A distribution network fault line selection method considering topology changes, characterized in that: include: S10: Data preprocessing: by collecting zero-sequence current data and converting it into graph structure data, the model is provided with input information characterizing the distribution network topology and current characteristics; S20: Model training: Use the graph attention network to learn the adjacency matrix and node features until the model can accurately predict the probability of single-phase grounding faults in the distribution network; S30: Fault judgment: Based on the node failure probability output by the model, a threshold is set to identify and locate the specific fault line; S40: Model application: Input the on-site zero-sequence current data into the trained model to quickly and accurately determine the fault location of the distribution network.

2. The method for fault line selection in distribution network considering topology structure change according to claim 1, characterized in that: The S10 includes: S101: Data collection: collecting zero-sequence current data from the distribution network site; S102: Data cleaning: removing noise and abnormal values ​​from the collected zero-sequence current data; S103: Feature extraction: extract key features from the cleaned data; S104: Constructing a graph structure: converting the node and branch information of the distribution network into graph structure data, where the nodes represent various parts of the distribution network and the edges represent the connection relationship between the nodes.

3. The method for fault line selection in distribution network considering topology structure change according to claim 2, characterized in that: The key features extracted from the data include current amplitude and / or phase angle.

4. The method for fault line selection in distribution network considering topology structure change according to claim 3 is characterized in that: The S104 includes: First, the topological connection relationship of the distribution network measurement nodes is obtained, and a node connection matrix is ​​made according to formula (1). The zero-sequence current data samples required for the experiment are obtained by measuring the data of the zero-sequence current transformer in the field. The sample data obtained from field measurements are normalized using formula (3) to normalize the fault current signal of each measuring point, and the processed experimental data are obtained, which are used as the node characteristic data of each node. The nodes on both sides of the fault section of the IEEE18-node network are regarded as fault nodes, and the label of the fault node is set to 1, and the label of the non-fault node is set to 0: Specifically: From the perspective of graph, without considering the internal components and line characteristics, the distribution network node network is abstracted as a graph structure data G = (X, A) of nodes and edges; the adjacency matrix A describes the connection relationship between the nodes of the distribution network, where the row and column numbers of the adjacency matrix correspond to the nodes, as shown in formula (1): （1）； The branch zero-sequence current is used as the characteristic vector of the branch head node; in the power system, the zero-sequence current refers to the current generated when the vector sum of the three-phase currents is not zero in a three-phase system; the zero-sequence current is usually related to ground faults; in actual scenarios, for cable lines, the zero-sequence current transformer is a device specifically used to detect zero-sequence current. It is installed at the neutral point of the three-phase power system and can detect the zero-sequence current generated by faults or unbalanced loads; the zero-sequence current transformer is used to collect zero-sequence current; and the current measurement of overhead lines usually uses three-phase current transformers, which are installed on each line of the three phases respectively; by measuring the three-phase current, the zero-sequence current can be calculated; the calculation method is as follows: （2）； in, , and They are the three-phase currents A, B, and C measured by the three-phase current transformer; X is the node feature vector: (3); is the number of nodes; Contains this node The sampling point time series samples are 1024 in this embodiment; the normalization method is used to preprocess the data, and the formula is as follows: （4）； in, is the normalized data, is the minimum value in the original data, is the maximum value in the original data; after this processing, the minimum value in the original data is mapped to 0, the maximum value is mapped to 1, and other values ​​are linearly mapped between 0 and 1.

5. The method for fault line selection in distribution network considering topology structure change according to claim 4, characterized in that: The S20 is specifically: The processed adjacency matrix and node features are input into the constructed graph neural network model. The spatial features and temporal features are extracted respectively through GAT and LSTM. Different weights are assigned to different nodes through the graph attention mechanism to aggregate more obvious spatial features. At the same time, the multi-gate structure unit of LSTM captures multiple dependencies of timing signals through forgetting, updating and outputting. The time and space feature vectors are spliced ​​together through a splicing layer. After passing through a fully connected layer and a sigmoid classification layer, the prediction results and the actual fault labels are calculated using the cross entropy loss function. According to the size of the loss value, the model parameters are updated through back propagation to optimize the model in the direction of smaller loss value. According to the probability Pi of each node failure output by the model, it is determined between which two nodes of the distribution network there is a fault, thereby determining the fault section. When the training result meets a certain accuracy, the model parameters are saved to obtain a trained model.

6. The method for fault line selection in distribution network considering topology structure change according to claim 5, characterized in that: The S20 includes: S201: Initialize the model: set the initial parameters of the graph attention network model, including weights and biases; S202: Input data: The adjacency matrix and node features of the constructed graph structure data are used as input data; S203: Forward propagation: Calculate the correlation between nodes through the graph attention mechanism and update the node features; S204: Loss function: Define a loss function to measure the difference between the model output and the actual fault location; S205: Back propagation: Calculate the gradient according to the loss function and update the model parameters; S206: Iterative optimization: repeat the forward propagation and back propagation process until the model loss reaches a preset threshold or the number of training times reaches a set value; S207: Model saving: saving the trained model parameters for subsequent application.

7. The method for fault line selection in distribution network considering topology changes according to claim 6, characterized in that: The S203 includes: The graph attention network structure is divided into two main parts, namely GAT and LSTM. Among them, GAT extracts spatial features and LSTM extracts temporal features. GAT introduces an attention mechanism based on GCN to set weights for the information aggregation of nodes and their adjacent nodes, thereby aggregating more significant features of each node. The input of each layer of GAT is a set of feature vectors of the node. ; N is the number of topological nodes in the distribution network; x i is the feature vector of node i; F is the feature number of the node; its output is a new feature set: Where x and Has different dimensions; First, calculate the attention coefficient between each node, expressed as （5）； where e ij is the attention coefficient of node j (neighboring node of node i) to node i; is the attention function, which maps the concatenated high-dimensional features to a real number; •||• is the feature concatenation operation; x i and x j is the input feature vector of node i and node j; W is the shared parameter matrix of linear transformation; N i is the total number of first-order neighbor nodes of node i; the calculated attention coefficient is nonlinearly transformed and normalized, as shown in the following formula （6）； After that, the adjacent node vectors are weighted and summed according to the attention weights to update the feature vector of the target node i, as shown in the following formula: （7）； Use a multi-head attention mechanism to improve the performance of the model, and set the number of attention heads to 4; that is, use 4 independent attention mechanisms to calculate the attention weights and update the node features, and then concatenate the vectors updated by each attention mechanism; LSTM is a variant of recurrent neural network, which makes up for the shortcomings of recurrent neural network that it cannot remember information from a long time ago and the gradient will explode and disappear. The LSTM hidden layer is formed by stacking multiple gate structures. Each gate structure contains an input gate, an output gate and a forget gate, as well as two functions sigmod and tanh. The calculation formula for each time step is as follows: （8）； In the formula is the output of the forget gate, , are the weights corresponding to the output of the previous gate structure and the input of the gate structure in the current period, is the output of the gate structure in the previous period, is the input of the gate structure of the current period, is the bias term; (sigmod) The value of each dimension output is between 0 and 1, 1 means that the original information is completely retained, and 0 means that the original information is completely removed; （9）； （10）； In formula (9) and (10) is the activation vector of the input gate, is the information accumulated at the current moment; multiply the two together to get the state quantity to be updated; （11）； In formula (11), The state input of the next gate structure is obtained by The state information that needs to be forgotten with the current structure Multiply and add the state amount that needs to be updated in the current state The state quantity of the next gate structure input can be obtained; （12）； （13）； Formula (12) is the output gate activation vector of LSTM, and formula (13) is the output of LSTM. It multiplies the activation vector of the output gate and the state of the current gate structure by tanh to get the output.

8. The method for fault line selection in distribution network considering topology structure change according to claim 7, characterized in that: The S30 includes: S301: Fault probability calculation: Use the trained model to predict the fault probability of the new graph structure data; S302: Fault location: According to the output node fault probability Pi, determine the section where the fault is most likely to occur; S303: Threshold setting: a fault probability threshold is set. Only when the fault probability of a node exceeds the threshold, the node is considered to be faulty.

9. The method for fault line selection in distribution network considering topology structure change according to claim 8, characterized in that: The S204 includes: The graph attention network aggregates the input adjacency matrix and node features through GAT for spatial feature aggregation, and uses LSTM in parallel with GAT to extract the time series features of zero-sequence current of each node; the feature vectors obtained by the two are spliced ​​so that each node has both the spatial features of the distribution network topology and the time series features of zero-sequence current. The features are nonlinearly transformed through two fully connected layers, and finally the probability p of each node being a faulty node is output through the Sigmoid activation function. i , thereby determining the fault section; the Sigmoid function formula is as follows: （14）； in is the input value, e is the base of natural logarithm; The output prediction result is compared with the original label, and the loss value is calculated using the Focal Loss loss function. Focal Loss adds an adjustment factor to reduce the weight of easy-to-classify samples on the basis of the balanced cross entropy loss function, focusing on the training of difficult samples. Its definition is as follows: （15）； in, is the probability value predicted by the model, is the balancing factor, is a regulating factor; and The value of can make Focal Loss pay more attention to difficult samples, thereby improving the performance of the model on unbalanced data sets.

10. The method for fault line selection in distribution network considering topology structure change according to claim 6, characterized in that: The S40 includes: S401: Data input: After preprocessing, the zero-sequence current data collected on site is input into the trained model; S402: Fault prediction: The model outputs the fault probability of each node and sorts them according to the probability; S403: Fault line determination: Determine the fault line in the distribution network according to the node with the highest fault probability; S404: Result output: output the fault line information.

Images