Power distribution network fault positioning method based on double-task learning

Abstract

The invention discloses a power distribution network fault positioning method based on double-task learning. The method comprises the following steps: firstly, collecting characteristic information of nodes and edges of the power distribution network, making a data set by taking a fault state and a fault type as double labels, and inputting the data set into a power distribution network single-phase earth fault positioning model after PCA dimension reduction processing; the power distribution network single-phase earth fault positioning model initializes a feature vector for each load node, converts an edge feature into a node feature, inputs the node feature into two graph convolutional neural networks connected through a cross-stitch unit after iterative convergence, further learns the node feature, and obtains a fault location result. And a parameter soft sharing method is adopted to train the two graph convolutional neural network models at the same time. And finally, outputting the fault state and type corresponding to the sample. According to the method, a double-task learning framework is combined with a power distribution network fault positioning theory, so that more detailed fault type analysis is provided while the fault is quickly positioned, and the method is suitable for the condition of topological change of the urban power distribution network.

Description

Technical Field

[0001] The present invention belongs to the technical field of power systems, relates to ground fault location of a distribution network, and specifically to a distribution network fault location method based on dual-task learning. Background Art

[0002] Urban distribution networks usually have complex feeder segment structures and diverse line configurations. Traditional fault location methods have certain limitations in dealing with these challenges. Since traditional measurement equipment cannot accurately collect sufficient data in real time, measurement errors are large, and data loss and distortion problems often occur. In addition, since the topology of the distribution network changes frequently due to factors such as equipment replacement, line maintenance, and fault isolation, traditional methods have poor adaptability. With the widespread access of distributed power sources and the improvement of users' requirements for power supply quality, the operating status and dynamic characteristics of urban distribution networks are becoming increasingly complex, resulting in more diverse and unpredictable safety hazards.

[0003] Traditional fault location methods, such as the impedance method, traveling wave method, and intelligent optimization algorithm based on heuristic search, have been successfully applied in small distribution networks, but they mainly rely on the measurement of data such as voltage and current, and require frequent matrix iteration and update, which is computationally intensive and the convergence speed of the location effect is slow in large-scale distribution networks. Therefore, when facing large-scale distribution networks, the efficiency and accuracy of these methods are relatively limited. Traditional methods are usually limited to the determination of the fault location and cannot accurately determine the fault type. This limits their application in complex distribution networks, because the determination of the fault type is crucial for timely response and effective repair.

[0004] In recent years, artificial intelligence-based methods, especially single-task research, have made some progress. These methods include node classification and regression models based on deep learning, which can accurately locate faulty nodes and provide auxiliary decision-making for fault repair. However, most existing methods focus on a single task, such as fault location or fault classification, and rarely involve multi-task learning methods that handle both fault location and fault classification tasks at the same time. Summary of the invention

[0005] In view of the shortcomings of the prior art, the present invention proposes a distribution network fault location method based on dual-task learning, which inputs the preprocessed distribution network feeder data into a graph convolutional neural network to locate the fault coordinates and type, thereby overcoming the shortcomings of traditional fault location methods that do not fully consider the distribution network topology and fault type judgment.

[0006] A distribution network fault location method based on dual-task learning specifically includes the following steps: Step 1: Use FTU equipment to collect current, voltage data and circuit breaker status on each branch line and load node on the distribution network, and integrate historical fault data as sample data. The distribution network status and fault type at the time of data collection are used as the dual labels corresponding to the sample data, and the features of the sample data are extracted to produce a training data set with dual labels.

[0007] Step 2: Build a single-phase grounding fault location model for distribution network based on graph convolutional neural network. The location model includes a feature conversion iteration module, a dual-task learning module and a feature output module.

[0008] The feature conversion iteration module is used to initialize the feature vector for each load node, convert the edge features of real-time data and historical data into node features and combine them, and repeat the conversion and aggregation process until the feature representation converges.

[0009] The input of the dual-task learning module is the feature representation output by the feature conversion iteration module. The node information errors are classified into two categories through two graph convolutional neural network models, and the feature vectors of node grounding fault and node grounding fault type are output respectively. A cross-stitch unit is used between the output of the hidden layer of the two graph convolutional neural network models and the input of the next hidden layer to connect the two graph convolutional neural network models, and the model is trained and optimized by using soft parameter sharing and minimizing the joint loss function to complete the dual-task learning of node grounding fault location and node grounding fault type.

[0010] The feature output module completes the transformation of node features into edge features according to the learning results output by the dual-task learning module, and outputs final fault location information, specifically no fault, single-phase grounding fault or non-single-phase grounding fault.

[0011] Step 3: Optimize the training of the single-phase grounding fault location model of the distribution network built in step 2; use the Adam optimizer and adopt the learning rate decay strategy to improve the training stability and effect; use AMP training to increase the running speed.

[0012] Step 4: Input the distribution network sample data with unknown status into the model optimized in step 3, output the node fault characteristics, and complete the fault location and fault classification.

[0013] The present invention has the following beneficial effects: Compared with the traditional fault location method, this method uses graph neural network to establish a dual-task learning model. Through the dual-task learning framework, the fault location task and the fault classification task share feature representation, but their optimization objectives are different. The fault location task can be used as a fast and accurate "fault warning" mechanism, while the fault classification task determines the fault type at a more complex level. In this way, the fault location task will not be covered by the fault classification task, but can provide faster detection results as an independent task. Compared with the traditional fault location method, this method can flexibly select the measurement data of current and voltage, has a certain fault tolerance for unknown data, can provide more detailed fault classification for the fault location task, and has better adaptability when the distribution network topology changes, which improves the topological generalization ability of the fault location model. Using the dual-task learning framework, the fault type is output while locating the fault location, providing more comprehensive information support for subsequent fault processing and isolation. This capability is particularly suitable for the increasingly complex operating conditions and dynamic characteristics in modern distribution networks. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 is a schematic diagram of an urban feeder group for fault location in an embodiment; Figure 2 This is a flow chart of a distribution network fault location method based on dual-task learning; Figure 3 is a schematic diagram of the dual-task learning module structure constructed in the embodiment; Figure 4 It is a comparison of the distribution network fault location accuracy training process of the dual-task model and the single-task model in the embodiment; Figure 5 It is a comparison of the distribution network fault classification accuracy training process of the dual-task model and the single-task model in the embodiment. DETAILED DESCRIPTION

[0015] The method is further explained below with reference to the accompanying drawings.

[0016] like Figure 1 As shown in the figure, each vertex of the S-SCDN (medium-voltage distribution network supply area centered on a high-voltage substation) topology diagram is an electrical node of the distribution network, and the edge is an overhead line or cable line. The entire S-SCDN feeder group network is a typical structure of a large number of generalized switching stations holding hands. There are many generalized switching stations inside, so there are many feeder branches and connecting lines, and there are a large number of segmented and partitioned circuit breakers.

[0017] like Figure 2 As shown, a distribution network fault location method based on dual-task learning specifically includes the following steps: Step 1: For Figure 1The CA transformer S-SCDN feeder group has 6 substation outgoing lines, 17 nodes, and 21 branches under it. An FTU is installed next to each circuit breaker for data collection.

[0018] Randomly select 50 data sampling points on each branch, including artificially set fault points and non-fault points. Each data point contains the positive, negative, zero sequence, amplitude, phase angle and position feature vector of the voltage and current as sample data. At the same time, integrate the historical fault data, take the distribution network state and fault type at the time of data collection as the corresponding double labels, and select the features to make a labeled data set. A total of 3000 sets of sample data were collected, including 1250 sets of single-phase grounding faults and non-single-phase grounding faults, and 500 sets of normal state. Randomly select 2500 sets as training set data, and the remaining 500 sets as test set data.

[0019] PCA is used to reduce the dimension of the collected current and voltage information to match the dimension of the model input: first, the sample data X is standardized, and then the covariance matrix is ​​calculated, the eigenvalues ​​and eigenvectors are solved, the contribution rate and cumulative contribution rate of each principal component are calculated, and the eigenvalues ​​and corresponding eigenvectors with cumulative contribution rates ≥ 95% are selected to form a matrix, and the characteristic matrix Y is obtained by transformation.

[0020] Step 2: Build a single-phase grounding fault location model for distribution network based on graph convolutional neural network, input the feature matrix Y obtained in step 1, obtain the corresponding fault state, and further determine the fault type. The location model includes a feature conversion iteration module, a dual-task learning module, and a feature output module.

[0021] The feature conversion iteration module is a single-layer fully connected neural network, which initializes a feature vector for each load node, converts the edge features of real-time data and historical data into node features and combines them, and repeats the conversion and aggregation process until the feature representation converges.

[0022] The input of the dual-task learning module is the feature representation output by the feature conversion iteration module, and the node grounding fault feature and the node grounding fault type feature are learned respectively. Figure 3 As shown in the figure, each graph convolutional neural network includes an input layer, six hidden layers and an output layer; the number of neurons in the input layer is 6; the hidden layer uses the GraphSAGE algorithm based on the LSTM aggregator, and the output feature dimensions of each hidden layer from input to output are 10, 10, 20, 20, 10, 10 respectively. The 1st to 5th hidden layers are graph convolutional layers and rectified linear units, which are used to consume and learn the features in the input feature matrix, and the last hidden layer is a convolutional layer.

[0023] In order to achieve dual-task learning, a cross-stitch unit is set between the output of each hidden layer of the two graph convolutional neural network models and the input of the next hidden layer, so as to connect the two graph convolutional neural network models. By learning the linear combination of the input features in the feature matrix, the representation sharing between faults and categories is modeled: in, They represent the outputs of the i-th hidden layer of the two graph convolutional neural network models, , They represent the input of the i+1th hidden layer respectively. i=1, 2, 3, 4, 5. The subscript F represents fault location, and D represents fault classification. and The same task parameters, that is, the part that corresponds to the parameterized input and then re-output to the original task. and The parameters of different tasks correspond to the part of the input parameterized and output to another task. Each task still retains its own optimization goal, avoiding one task completely covering the unique characteristics of another task, suppressing overfitting, and enhancing the robustness of the model.

[0024] One of the graph convolutional neural network models extracts node feature information layer by layer through 6 hidden layers, aggregates and updates the topological relationship and feature data between the node and its neighboring nodes multiple times, and finally obtains the deep feature representation of each node for subsequent fault location. The other graph convolutional neural network model further extracts high-order features based on node features through 6 hidden layers, and combines the topological structure and existing feature information between nodes to provide support for subsequent fault type judgment. The dual-task learning module finally outputs two sets of feature vectors based on node features. This design realizes feature sharing and independent optimization between tasks, laying the foundation for subsequent accurate classification.

[0025] The goal of fault location in the distribution network is to identify the faulty line, so it is necessary to generate edge features based on node features and perform edge classification. The feature output module includes a fully connected layer and a Softmax function, which is used to generate edge features from two sets of feature vectors based on node features output by the dual-task learning module, and output the final discrimination result. Specifically, for each edge, the 10-dimensional feature vectors of the two nodes connected to it are synthesized into a 20-dimensional feature vector through a splicing operation as the input of the fully connected layer. Through the mapping of the fully connected layer, the 20-dimensional feature vector is converted into a 2-dimensional edge feature vector. Subsequently, the Softmax function normalizes the output of the fully connected layer to generate a classification probability distribution for each edge, and outputs the results of fault location and fault classification respectively. Among them, the fault status discrimination results of the output nodes and branches are , the result of fault type determination , since both tasks are designed as binary classification problems, and the accuracy of task 1 alone is higher, The output is set to 0 or 1. is the binary probability output. Multiplying the two output result vectors will give the final fault feature T, which contains three elements: Indicates no fault. represents the probability of single-phase grounding fault, Represents the probability of non-single-phase grounding fault.

[0026] Step 3: Optimize the training of the single-phase grounding fault location model of the distribution network built in step 2, and design the following joint loss function: : , They represent the loss function of the fault location task and the loss function of the fault classification task respectively, W is the model weight, and For the parameters that need to be adjusted, the model parameters are balanced by minimizing the joint loss function.

[0027] Use AMP training to increase running speed; use Adam optimizer and adopt learning rate decay strategy to improve training stability and effect: epoch represents the training round. Represents the learning rate.

[0028] Step 4: For the distribution network sample data with unknown status, extract the feature matrix Y by the method of step 1, input the model optimized in step 3, and obtain the node fault characteristics, specifically no fault, single-phase grounding fault or non-single-phase grounding fault.

[0029] This embodiment uses 3000 sets of sample data, including 1250 sets of single-phase grounding faults, 1250 sets of non-single-phase grounding faults and 500 sets of normal state samples. In order to train the model, 2500 sets of samples are randomly selected for model training, and the remaining 500 sets of samples are used to test the performance of the trained model. Figure 4The comparison results of the dual-task learning method and the traditional single-task method in the fault location task are shown. The experimental results show that after 350 rounds of training, the accuracy of the dual-task learning model in the fault location task is 97.2%, while the accuracy of the single-task model fluctuates greatly after reaching a certain level, but generally tends to be stable, with a final accuracy of 90.6%. This result shows that when the traditional single-task method faces complex distribution network topology, the model is not stable enough during the iteration process, the fault location accuracy is limited, and it is difficult to fully capture the correlation information between nodes. In contrast, the dual-task learning method can effectively share node features and topology information by simultaneously optimizing the fault location and fault classification tasks, thereby improving the accuracy of fault location.

[0030] Figure 5 The comparison results of the dual-task learning method and the traditional single-task method on the fault classification task are shown. Through comparative analysis, it can be seen that the accuracy of the dual-task learning model on the fault classification task is 81.6%, while the fault classification accuracy of the single-task model is only 65.2%. The main reason for this difference is that the traditional single-task method usually only focuses on fault classification and fails to fully utilize the auxiliary information provided by the fault location task. In contrast, the dual-task learning framework significantly enhances the judgment ability of the classification model in complex situations by sharing the feature information of the fault location task, verifying the significant performance advantage of this method compared to the traditional single-task processing method.

[0031] The above examples only express the implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for ordinary technicians in this field, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the patent of the present application shall be subject to the attached claims.

Claims

1. A distribution network fault location method based on dual-task learning, characterized by: The specific steps include: Step 1: Use FTU equipment to collect current, voltage data and circuit breaker status on each branch line and load node on the distribution network, and integrate historical fault data as sample data; use the distribution network status and fault type at the time of data collection as the dual labels corresponding to the sample data, extract the features of the sample data and make a training data set with dual labels; Step 2: Build a single-phase grounding fault location model for distribution network based on graph convolutional neural network, the location model includes a feature conversion iteration module, a dual-task learning module and a feature output module; The feature conversion iteration module is used to initialize the feature vector for each load node, convert the edge features of real-time data and historical data into node features and combine them, and repeat the conversion and aggregation process until the feature representation converges; The input of the dual-task learning module is the feature representation output by the feature conversion iteration module, and the node information errors are classified into two categories through two graph convolutional neural network models, and the feature vectors of the node grounding fault and the node grounding fault type are output respectively; the output of the hidden layer of the two graph convolutional neural network models is connected with the input of the next hidden layer using a cross-stitch unit, and the model is trained and optimized by using parameter soft sharing and minimizing the joint loss function to complete the dual-task learning of the node grounding fault location and the node grounding fault type; The feature output module completes the transformation of node features into edge features according to the learning results output by the dual-task learning module, and outputs final fault location information, specifically, no fault, single-phase grounding fault or non-single-phase grounding fault; Step 3: Optimize and train the distribution network single-phase grounding fault location model built in step 2; Step 4: Input the distribution network sample data with unknown status into the model optimized in step 3, output the node fault characteristics, and complete the fault location and fault classification.

2. A distribution network fault location method based on dual-task learning as claimed in claim 1, characterized in that: The sample data includes positive, negative, zero sequence, amplitude, phase angle of voltage and current and position characteristic vector of circuit breaker.

3. A distribution network fault location method based on dual-task learning as claimed in claim 1 or 2, characterized in that: The PCA dimension reduction is performed on the sample data, and the eigenvalues ​​and corresponding eigenvectors with cumulative contribution rates ≥ 95% are selected to form a matrix, which is input into the single-phase grounding fault location model of the distribution network.

4. A distribution network fault location method based on dual-task learning as claimed in claim 1, characterized in that: The feature conversion iteration module is a single-layer fully connected neural network; The feature output module includes a fully connected layer and a Softmax function.

5. A distribution network fault location method based on dual-task learning as claimed in claim 1, characterized in that: The dual-task learning module learns node grounding fault features and node grounding fault type features respectively through two graph convolutional neural network models; each graph convolutional neural network includes an input layer, six hidden layers and an output layer; the number of neurons in the input layer is 6; the hidden layer adopts the GraphSAGE algorithm based on the LSTM aggregator, and the output feature dimensions of each hidden layer from input to output are 10, 10, 20, 20, 10, 10 respectively; the 1st to 5th hidden layers are graph convolutional layers and rectified linear units, which are used to consume and learn the features in the input feature matrix, and the last hidden layer is a convolutional layer.

6. A distribution network fault location method based on dual-task learning as claimed in claim 1 or 5, characterized in that: The cross-stitch unit learns linear combinations of the input features in the feature matrix to model the representation sharing between two graph convolutional neural networks: in, They represent the outputs of the i-th hidden layer of the two graph convolutional neural network models, , They represent the input of the i+1th hidden layer respectively; i=1, 2, 3, 4, 5; the subscript F represents fault location, and D represents fault classification; and For the same task parameters, and Parameters for different tasks.

7. A distribution network fault location method based on dual-task learning as claimed in claim 1, characterized in that: Designing a joint loss function Optimize the training of the single-phase grounding fault location model of the distribution network: , They represent the loss function of the fault location task and the loss function of the fault classification task respectively, W is the model weight, and For the parameters that need to be adjusted, the model parameters are balanced by minimizing the joint loss function; Use AMP training to increase running speed; use Adam optimizer and adopt learning rate decay strategy to improve training stability and effect: epoch represents the training round. Represents the learning rate.

Images