A Single-Phase Grounding Fault Location Method for Distribution Networks Based on Graph Convolutional Neural Networks

Abstract

The present invention discloses a single-phase grounding fault location method for a distribution network based on a graph convolutional neural network. The method includes: collecting the characteristic information of load nodes and edges in different states of the distribution network feeder section through FTU terminal devices, normalizing the data, and then using a single-layer fully connected neural network to convert the edge feature data into node feature data connected thereto. Then, all node feature numbers are connected into a matrix and input into the graph convolutional neural network model, and the state of the distribution network during data collection is used as a label to train the network to obtain the structural parameters of the optimal graph convolutional neural network. Finally, based on the trained location model, faults occurring at each feeder of the substation are searched and located, so that the specific branch where the fault is located can be obtained through the output feature data. The present invention uses a neural network method to integrate the distribution network fault location theory, which can effectively ensure the location accuracy of faults in complex urban distribution networks and is suitable for the situation under the topological changes of urban distribution networks.

Description

Technical Field

[0001] The present invention belongs to the technical field of power systems, relates to the grounding fault location of a distribution network, and specifically relates to a single-phase grounding fault location method for a distribution network based on a graph convolutional neural network. Background Art

[0002] The internal feeder segments of an urban distribution network have many branches and a complex line structure. For single-phase grounding faults, if traditional fault location methods are used, due to the inability of traditional measurement devices to collect enough real-time and accurate information, there are large measurement errors, and data loss and distortion are likely to occur. In addition, due to reasons such as equipment replacement, line maintenance, and fault isolation, the network topology of the distribution network changes relatively frequently. With the large-scale integration of distributed power sources such as wind power, photovoltaic power generation, energy storage devices, and microgrids, the change of the network structure is more uncertain. Traditional fault location methods mainly based on matrix methods and optimization methods such as genetic algorithms and particle swarm algorithms cannot adapt to the urban distribution network structure with large topological feature changes because they often need to perform iterative updates of matrices, have a high dependence on data such as voltage and current, and the calculation amount increases. Summary of the Invention

[0003] Aiming at the deficiencies of the prior art, the present invention proposes a single-phase grounding fault location method for a distribution network based on a graph convolutional neural network. After processing the original distribution network feeder data, it is input into the graph convolutional neural network for single-phase grounding fault location, overcoming the shortcoming that traditional fault location methods do not fully consider the distribution network topology.

[0004] A single-phase grounding fault location method for a distribution network based on a graph convolutional neural network specifically includes the following steps:

[0005] Step 1: Use FTU (Feeder Terminal Unit) devices to collect the current, voltage signals, and breaker data of each branch line and load node on the distribution network as sample data, and then use the distribution network state during collection as the corresponding sample label, including fault state and non-fault state, so as to obtain a labeled data set. Perform a normalization operation on the sample data in the data set to enhance the generalization ability of the samples.

[0006] Step 2: Build a single-phase grounding fault location model for a distribution network based on a graph convolutional neural network. The location model includes a feature conversion module, a feature extraction module, and a feature output module.

[0007] The feature conversion module includes a single-layer fully connected neural network (MLP) for converting the current and breaker feature data of the edges in the data set obtained in Step 1 into the feature data of the load nodes connected thereto, and then inputting it into the feature extraction module.

[0008] The feature extraction module is used to input the load node feature data converted by the feature conversion module into the graph convolutional neural network to extract the feature data of each load node in the distribution network topology.

[0009] The feature output module, including a fully connected neural network layer and a Softmax function, is used to output the feature vectors of node ground faults and line ground faults according to the feature data output by the feature extraction module.

[0010] Step 3: Construct a model optimization module to optimize and train the positioning model built in Step 2. The model optimization module includes an Adam optimizer, a Dropout layer, and a pooling layer based on LSTM. Among them, the Dropout layer is introduced before the convolutional layer of the graph convolutional neural network, and the pooling layer based on LSTM is introduced after the convolutional layer of the graph convolutional neural network. The Adam optimizer is used for the optimization and iterative training of the positioning model.

[0011] Step 4: Input the distribution network sample data with unknown status into the positioning model optimized in Step 3 to obtain the feature vectors of node ground faults and line ground faults for fault location.

[0012] The present invention has the following beneficial effects:

[0013] Compared with the traditional fault location method, this method uses a graph neural network to establish a single-phase ground fault location model in the distribution network. Since the GCN introduces the message passing and aggregation mechanism of nodes, compared with the traditional fault location method, this method can have a certain tolerance for unknown data, reduce the dependence on current and voltage measurement data, and has better adaptability when the distribution network topology changes, improving the topology generalization ability of the fault location model. Brief Description of the Drawings

[0014] Figure 1 is a schematic diagram of the urban feeder group for fault location in the embodiment;

[0015] Figure 2 is a flowchart of the single-phase ground fault location method for the distribution network;

[0016] Figure 3 is a flowchart of the "edge-node" feature conversion in the embodiment;

[0017] Figure 4 is a schematic diagram of the single-phase ground fault location model for the distribution network constructed in the embodiment.

[0018] Figure 5 is an F1-Score curve graph of the model based on the test set under 300 iterations in the embodiment. Detailed Embodiments

[0019] Taking S-SCDN (the medium-voltage distribution network supply area centered on a high-voltage substation) as an example, the present invention will be further explained in combination with Figure 1 the urban feeder group shown in the figure; each vertex of the S-SCDN topology diagram is an electrical node of the distribution network, and the edges are overhead lines or cable lines. The entire S-SCDN feeder group network is a typical structure in the form of a large number of generalized switchgear (switching stations, ring stations, etc.) connected in a hand-in-hand manner. There are many generalized switchgear inside, so there are many feeder branches and tie lines, and there are a large number of sectional and sectional circuit breakers.

[0020] As Figure 2 shown, a single-phase grounding fault location method for a distribution network based on a graph convolutional neural network specifically includes the following steps:

[0021] Step 1: For the 2 feeders, 12 nodes, and 15 branches under the Xiangqiao Substation, and the 2 feeders, 7 nodes, and 8 branches under the Chengguan Substation, install FTUs beside each circuit breaker for measurement and acquisition. Randomly take 100 sampling data points on each branch, including data of artificially set fault points and non-fault points. Each data point includes node voltage, current, and the circuit breaker position feature vector of the edge. A total of 2000 groups of sample data are collected and divided into a training set, a validation set, and a test set according to a ratio of 5:3:2. Among them, the training set includes 1000 samples, the validation set includes 600 samples, and the test set includes 400 samples.

[0022] In order to eliminate the adverse effects caused by singular feature data and facilitate subsequent data reading work of the program, the node feature data is encoded in a one-hot manner, and the feature data is standardized, and the input feature size is limited within [0,1]

[0023]

[0024] where x represents the feature vector of the sample, x' is the feature vector after standardization, max(x) is the maximum value of the samples in the dataset, and min(x) is the minimum value of the samples in the dataset.

[0025] Step 2: Build a single-phase grounding fault location model for a distribution network based on a graph convolutional neural network. The location model includes a feature transformation module, a feature extraction module, and a feature output module. In this embodiment, the training and testing of the network model are both carried out in the Python 3.6.3 environment, and the Pytorch framework of version 1.10.2 is used. The network model is built using the corresponding versions of the cluster, gemetric, scatter, sparse, and spline-conv modules, and Tensorflow 2.6.2 is used for calculation implementation.

[0026] The feature transformation module includes a single-layer fully connected neural network (MLP), which receives the feature vectors standardized in step 1 and converts the current and circuit breaker feature data of the edges into the feature data of the connected load nodes, realizing the "edge-node" feature transformation. As Figure 3 shown, the specific feature transformation process is as follows:

[0027] s2.1. For a node P with L adjacent edges, the feature vector of its l-th adjacent edge is E l . According to the dimension d of E l , a single-layer fully connected neural network with d inputs and d outputs is constructed, and a bias term b is added.

[0028] s2.2. The feature vectors E1 to E l of the L adjacent edges connected to node P are all input into the fully connected neural network constructed in s2.1, and the outputs E1' to E l ' with the same dimension are obtained.

[0029] s2.3. The vectors E1' to E l ' obtained in s2.2 are summed vectorially to obtain a d-dimensional sum vector E sum , E sum = E′1 ∪ … ∪ E′ L .

[0030] s2.4. The feature vector T of node P and the sum vector E sum are combined into a new vector through a concatenation operation as the transformed feature vector T′, T′ = E sum + T.

[0031] During the fault location process, for each load node in the distribution network, its voltage loss fault signal needs to be considered. For the adjacent edges of the load node, its overcurrent fault signal and the circuit breaker position signal before the fault need to be considered. The feature transformation module converts the "edge features" into "node features", and the transformed feature vector T′ = X n = {x ni}, i ∈ [1, 4], x n1 to x n3 represent current, voltage, and circuit breaker position respectively, and x n4 is used as a label to reflect the state during data acquisition. When x n4 = 0, it represents the normal state, and when x n4 = 1, it represents the fault state. Each sample feature after transformation is concatenated into a matrix X and input into the feature extraction module, X = {X1, X2, …, X n , …, X N} T , X ∈ R N×4 , where N represents the number of nodes in the distribution network, n ∈ [1, N], Xn Represents the nth node feature after transformation.

[0032] The feature extraction module is used to input the load node feature data converted by the feature conversion module into the graph convolutional neural network to extract the feature data of each load node in the distribution network topology. As Figure 4 shown, in this embodiment, the graph convolutional neural network includes 1 input layer and 8 graph convolutional layers. The number of neurons in the input layer is 4, which is used to receive the 4-dimensional feature vector converted by the feature conversion module. The graph convolutional layer adopts the GraphSAGE algorithm based on the LSTM aggregator, which is an 8-layer hidden layer GraphSAGE neural network. Each graph convolutional layer obtains new node features through the aggregation operation between nodes, and then performs feature sampling and message passing through the GraphSAGE graph convolutional layer. The specific process is as follows: ① Each node samples the features of its adjacent nodes; ② Each node aggregates the information of neighbor nodes to update its own node features; ③ According to the updated node feature information, the model is trained. The GraphSAGE algorithm can overcome the limitation of the explosion of data volume during the training of GCN, thus ensuring that the model is applicable to the situation of changes in the distribution network topology.

[0033] The feature output module includes a fully connected neural network and a Softmax function, which are used to realize the "node-edge" feature conversion according to the feature data output by the feature extraction module, and output the state discrimination results Y of the nodes and branches n ={y ni}, i ∈ [1, 2], y n1 Node grounding fault state, y n2 Line segment grounding fault state.

[0034] Step 3: Build a model optimization module to optimize and train the positioning model built in Step 2. The model optimization module includes an Adam optimizer, a Dropout layer, and an LSTM-based pooling layer. The Dropout layer is introduced before the convolutional layer of the graph convolutional neural network to prevent overfitting during the learning process. The LSTM-based pooling layer is introduced after the convolutional layer of the graph convolutional neural network. Input the training set and validation set samples, and use the Relu activation function, cross-entropy error loss function, and Adam optimizer to perform 300 times of optimization and iterative training on the positioning model. The Adam optimizer can update the variables according to the oscillation situation of the historical gradients and the real historical gradients after filtering the oscillations.

[0035] Step 4: Input the sample data in the test set into the optimized positioning model in Step 3 to obtain the feature vectors of node grounding faults and line grounding faults for fault location.

[0036] For the fault location results in step 4, the F1-Score is used to evaluate the effectiveness of the location model:

[0037]

[0038] Among them, TP represents the number of positive samples that are predicted as positive samples, FP represents the number of negative samples that are wrongly predicted as positive samples by the model, and FN represents the number of positive samples that are wrongly predicted as negative samples by the model. As the harmonic mean of the precision_rate and recall_rate, the F1-score takes both of these two indicators into account. It can be considered that the larger the value of F1, the more accurate the fault discrimination. Since the number of iterations set during model training is 300 times, the F1-Scrore curve of this method is as shown in Figure 5 shown.

[0039] From Figure 5 it can be seen that as the number of iterations increases, the F1-Scrore value of this GCN model on the test set is getting higher and higher, approaching 1. Therefore, it can be seen that the overall fault recognition accuracy of the model has good performance and certain stability.

[0040] The fault location model based on GCN established by this method has good practical application performance. Based on S-SCDN, it can be applied to the topological changes in the scenario of distribution network reconstruction in urban distribution networks. Subsequently, deeper research can be carried out for different distribution networks to establish a general and transferable distribution network fault location model, and further improve the generalization and unknown capabilities of the topological model.

[0041] The above-described examples only represent the implementation modes of this application. The description is relatively specific and detailed, but it should not be construed as a limitation on the scope of the invention patent. It should be noted that for those of ordinary skill in the art, without departing from the concept of this application, several modifications and improvements can still be made, and these all belong to the protection scope of this application. Therefore, the protection scope of the patent of this application should be subject to the appended claims.

Claims

1. A single-phase grounding fault location method for a distribution network based on a graph convolutional neural network, characterized in that: The method specifically includes the following steps: Step 1: Use the FTU device to collect the current, voltage signals and breaker data of each branch line and load node on the distribution network as sample data, and then use the state of the distribution network during collection as the corresponding sample label to make a labeled data set; Step 2: Build a single-phase grounding fault location model for the distribution network based on a graph convolutional neural network. The location model includes a feature conversion module, a feature extraction module, and a feature output module; The feature conversion module includes a single-layer fully connected neural network, which is used to convert the current and breaker feature data of the edge into the feature data of the load node connected to it, and then input it into the feature extraction module; The feature extraction module is used to input the load node feature data converted by the feature conversion module into the graph convolutional neural network to extract the feature data of each load node in the distribution network topology; The feature output module includes a layer of fully connected neural network and a Softmax function, which is used to output the feature vectors of node grounding faults and line grounding faults according to the feature data output by the feature extraction module; Step 3: Build a model optimization module. After normalizing the data set made in Step 1, input it to optimize and train the location model built in Step 2. The model optimization module includes an Adam optimizer, a Dropout layer, and a pooling layer based on LSTM. The Dropout layer is introduced before the convolutional layer of the graph convolutional neural network, and the pooling layer based on LSTM is introduced after the convolutional layer of the graph convolutional neural network. The Adam optimizer is used for the optimization and iterative training of the location model; Step 4: Input the sample data of the distribution network with unknown status into the optimized location model in Step 3 to obtain the feature vectors Y of the node grounding fault and the line grounding fault n ={y ni}, i ∈ [1, 2], where y n1 represents the node grounding fault state, and y n2 represents the line segment grounding fault state.

2. The single-phase grounding fault location method for a distribution network based on a graph convolutional neural network according to claim 1, wherein: The normalization operation is as follows: Where x represents the feature vector of the sample, x′ is the feature vector after normalization, and x′∈[0,1], max(x) is the maximum value of the samples in the data set, and min(x) is the minimum value of the samples in the data set.

3. The single-phase grounding fault location method for a distribution network based on a graph convolutional neural network according to claim 1, wherein: The specific process of feature conversion by the feature conversion module is: s2.

1. For a node p with L adjacent edges, the feature vector of its l-th adjacent edge is E l ; Based on the dimension d of E l , construct a single-layer fully connected neural network with d inputs and d outputs, and add a bias term b for feature transformation; s2.

2. Input the feature vectors E1 to E of the L adjacent edges connected to node P into the fully connected neural network constructed in s2.1 to obtain the outputs E1' to E of the same dimension l '; l ' S2.

3. Perform vector summation on E1’~E l ’ obtained in S2.2 to obtain a sum vector E of dimension d sum , E sum = E′1 ∪…∪ E′ L ; S2.

4. Combine the feature vector T of node P and the sum vector E sum through a concatenation operation into a new vector as the transformed feature vector T′, where T′ = E sum + T.

4. The single-phase grounding fault location method for a distribution network based on a graph convolutional neural network according to claim 3, wherein: The transformed feature vector T' = X n = {x ni}, i ∈ [1, 4], where x n1 to x n3 represent current, voltage, and circuit breaker position respectively, and x n4 serves as a label reflecting the state during data collection. When x n4 = 0, it indicates a normal state, and when x n4 = 1, it indicates a faulty state.

5. The single-phase grounding fault location method for a distribution network based on a graph convolutional neural network according to any one of claims 1, 3, or 4, characterized in that: The feature transformation module concatenates each transformed sample feature into a matrix X and inputs it into the feature extraction module, X = {X1, X2, …, X n , …, X N}, T , X ∈ R N×4 , where N represents the number of distribution network nodes, n ∈ [1, N], and X n represents the nth node feature after transformation.

6. The single-phase grounding fault location method for a distribution network based on a graph convolutional neural network according to claim 1, wherein: The graph convolutional neural network includes 1 input layer and 8 graph convolutional layers. The number of neurons in the input layer is 4, which is used to receive the features converted by the feature conversion module. The graph convolutional layer uses the GraphSAGE algorithm based on the LSTM aggregator.

Images