Active distribution network topology identification method and related device based on multi-task graph learning

Abstract

The present invention discloses a method and related device for topology identification of an active distribution network based on multi-task graph learning, belonging to the technical field of power system monitoring, and the method comprises using a plurality of autoencoders to respectively complete the missing of different measurement features in the distribution network; representing the topology of the distribution network in the form of a graph, training the topology identifier using graph data, performing message aggregation from the target node to the source node according to the multi-head attention coefficient, and obtaining the feature matrix of the whole graph edge; inputting the feature of each line into a fully connected neural network layer, and mapping the feature into the judgment value of the state of each line, judging the disconnection state of the line, and realizing topology identification; using multi-task learning to jointly learn and optimize the missing of the complement measurement feature and the topology identification process, determining the target model, inputting the real-time measurement data into the target model, and outputting the topology identification result. The present invention can reduce the dimension, reconstruct the incomplete data, and accurately identify the topology structure of the distribution network under the condition of limited measurement data.

Description

Technical Field

[0001] The present invention belongs to the technical field of power system monitoring, and specifically relates to an active distribution network topology identification method based on multi-task graph learning and related devices. Background Art

[0002] The distribution network is a key component of the power system. As the penetration rate of renewable energy in the distribution network continues to increase, it is necessary to frequently reconstruct the distribution network topology to cope with the uncertainty brought by the access of distributed energy. The distribution network management system relies on real-time monitoring data, especially accurate topology information, to grasp the system operation status and ensure reliable operation. However, in practice, compared with the redundant transmission network, the distribution network is limited by factors such as cost, communication channel capacity and equipment failure. The measurement terminal is insufficiently configured and the measurement data is often delayed or missing. Traditional identification methods that rely on complete and accurate measurement data often find it difficult to carry out real-time topology tracking. Therefore, the distribution system urgently needs to realize online topology identification efficiently and accurately under limited measurement conditions to support the online transparency of the situation and ensure the power supply reliability of the system.

[0003] The current topology identification methods are mainly divided into traditional methods and artificial intelligence-based identification methods. Traditional identification methods are limited in processing the bidirectional power flow caused by the access of distributed power sources, and when the distribution network has a ring network structure, it is difficult to meet the needs of accurate identification of the distribution network topology. In addition, these traditional methods have problems of low efficiency and poor timeliness when processing large amounts of calculations. Another type of artificial intelligence-based identification method achieves identification by mining the correlation between measurement information and topology. Compared with traditional methods, it has stronger adaptability and can realize online real-time identification of topology. However, existing topology identification methods all need to rely on complete measurement data to achieve accurate identification results. In the face of missing measurements, their adaptability and effectiveness need to be improved, which leads to problems such as low economy and poor feasibility of practical applications. Summary of the invention

[0004] The purpose of the present invention is to provide an active distribution network topology identification method and related devices based on multi-task graph learning to address the problems in the above-mentioned prior art, so as to realize online transparency of the distribution network situation under limited measurement conditions, reduce the dimension, reconstruct incomplete data and accurately identify the distribution network topology structure under the conditions of limited measurement data.

[0005] In order to achieve the above object, the present invention has the following technical solutions:

[0006] In a first aspect, a method for identifying active distribution network topology based on multi-task graph learning is provided, comprising:

[0007] Use multiple pre-built autoencoders to complete the missing of different measurement features in the distribution network, and obtain the distribution network topology after the missing measurement features are completed;

[0008] The distribution network topology after missing measurement features are completed is represented as a graph to obtain data in the form of a graph;

[0009] The pre-built topology identifier is trained using graph data, and messages are aggregated from the target node to the source node according to the multi-head attention coefficient of the graph. The source node is the node currently being calculated, and each target node is the other node in the graph. The node association matrix is ​​used to calculate the difference in the feature vectors of the head and tail nodes after the update, and the features of the head and tail nodes are concatenated to form the head and tail node feature vectors, and the two-dimensional features of the line itself are obtained. The feature vector difference of the head and tail nodes after the update is concatenated with the head and tail node feature vectors and the two-dimensional features of the line itself to form the final feature matrix of the edge of the entire graph.

[0010] According to the final feature matrix of the whole graph edge, the features of each line are obtained, and the features of each line are input into the fully connected neural network layer, and the features are mapped to the judgment value of each line state. By setting the threshold of the judgment value for comparison, the disconnection state of the line is determined to achieve topology identification;

[0011] Multi-task learning is used to supplement the missing measurement features in the distribution network and optimize the joint learning of the topology identification process. The multi-task learning target model is determined, the real-time measurement data is input into the multi-task learning target model, and the topology identification results are output.

[0012] As a preferred solution, the step of using a plurality of pre-built autoencoders to respectively complete the missing of different measurement features in the distribution network includes:

[0013] In the encoding stage, the encoder fills the missing values ​​in the input measurement data with 0, extracts and captures the key information of the input features through the fully connected neural network layer, and outputs the encoded compressed representation. The expression is as follows:

[0014]

[0015] In the formula, for l The output vector of the layer encoder represents the dense representation of the input measurement data of the distribution network; , The encoder is Trainable weight parameters and biases of the layer; For the l- The output vector of the 1-layer encoder, is the selected parameterized ReLU activation function;

[0016] In the decoding stage, the decoder receives the compressed representation generated by the encoder and generates a completion measurement similar to the original input through a neural network layer that is reversed from the fully connected neural network layer of the encoder. The expression is as follows:

[0017]

[0018] In the formula, For decoder The output vector of the layer, i.e. the reconstructed missing values; , Decoder No. Trainable weight parameters and biases of the layer; For decoder The output vector of the layer.

[0019] As a preferred solution, when the distribution network topology after the missing measurement features are completed is represented in the form of a graph, the form of the graph is , where is the set of distribution network nodes, Indicates the number of electrical nodes in the distribution network, is the set of edges of the graph, consisting of overhead lines or cable lines, using Represents the adjacency matrix of the graph; distribution network topology after missing measurement features are completed, where node voltage amplitude, node active power, and node reactive power are modeled as node features in the graph. Each node represents a bus in the power system, and the node feature is a three-dimensional vector , the current amplitude and branch active power are used as edge features in the graph, each edge represents a line in the power system, and the edge feature is a two-dimensional vector ; The initial feature matrix of the graph is , where Is a node The initial attribute information vector, Is A vector composed of initial attributes; for a graph data that needs to classify edges, the true labels of the edge states are constructed as a label matrix, expressed as ,in, is the number of categories for classification, consisting of the one-hot encoding of the categories.

[0020] As a preferred solution, the topology identifier pre-built is trained using graph data, and in the step of aggregating messages from the target node to the source node according to the multi-head attention coefficient of the graph, the topology identifier is built based on the Graph Transformer algorithm, and the node features in the initial graph data are After algorithm aggregation, the updated attribute information of all nodes is obtained ; The measured characteristics of a given node , corresponding to the initial node features in the initial graph data, is calculated from each target node as follows To the source node The multi-head attention coefficient corresponding to each edge of :

[0021]

[0022]

[0023]

[0024]

[0025] In the formula, for the Attention head, l Layer-by-layer training weight matrices and the bias vector The node i Features Encoded as a query vector , training weight matrix and the bias vector The target node j Features Encoding key vector ; and train the weight matrix and the bias vector Edge Features Encoded as a relation vector Then added to the key vector as additional information for each layer; For source node and the target node The normalized attention weight coefficient between The numerator is the source node and the target node The attention value between the source nodes is The sum of the attention values ​​between all nodes in ; is the exponentially scaled dot product function, d is the hidden size of each attention head, is the number of nodes, and the query vector to the source node is obtained by point multiplication and the target node information;

[0026] According to the obtained multi-head attention coefficient, the multi-head attention matrix is ​​used as the transfer matrix of message passing, from the target node To the source node For message aggregation, the calculation expression is as follows:

[0027]

[0028]

[0029] In the formula, the weight matrix and the bias vector The target node Features Encoded as a vector of values After that, the relationship vector containing edge feature information Superposition, and then through the attention weight coefficient After weighted selection and aggregation, the next layer of source nodes is obtained Attention characteristics ,in Represents a splicing operation, is the number of attention heads, For source node Number.

[0030] As a preferred solution, it also includes using gated residual connections between the layers of the topology identifier to improve the layer structure. The calculation expression is as follows:

[0031]

[0032]

[0033]

[0034] In the formula, through the weight matrix and the bias vector The source node exist l Layer feature information Encoded as reset gating vector After that, with the source node exist l +1 layer of attention features The difference is encoded to obtain the updated gating vector ,in Represents the weight matrix Respectively , The weighted product is performed with the difference between the two, and sigmoid is used as the activation function; finally, the improved source node is obtained through the weighted product New Features ,in Representation layer normalization, For source node In updating the gating vector The strength of the corresponding residual connection in Simultaneous use l Layer features and l +1 layer output, increasing the diversity of features between different layers;

[0035] It also includes removing nonlinear transformation from the output and taking an average approach to obtain a final node representation for topology identification.

[0036] As a preferred solution, the final feature matrix of the entire graph edge is ,in , is the feature vector difference of the head and tail nodes, the feature vector of the head and tail nodes and Concatenate to form the head and tail node feature vectors , and They are the original features of the line and the features after completion;

[0037] The Sigmoid function is used to map the features into the judgment value of each line state. , the judgment value The set threshold is compared, and if it is greater than the threshold, it is considered to be 1, otherwise it is considered to be 0, where 1 represents that the circuit breaker is in a closed state, and 0 represents that the circuit breaker is in an open state.

[0038] As a preferred solution, the loss function expression of the multi-task learning target model is as follows:

[0039]

[0040] In the formula, , Indicates the first and The loss function of each task; , Represents subtasks The first predicted value and the true value; , Represents subtasks No. i Predicted values ​​and true values; Represents the linear combination of the loss functions of each subtask;

[0041] When multi-task learning is used to optimize the learning of missing measurement features in the distribution network, the loss function is established with the mean square error MSE between the completed value and the true value as the object, and the calculation expression is as follows:

[0042]

[0043] In the formula, is the mean square error (MSE) value of measurement completion; is the number of samples; For sample The true value of For sample The complement value of ;

[0044] Different autoencoders are used for different measurement types, and the loss functions of each measurement are instantiated as follows:

[0045]

[0046]

[0047]

[0048]

[0049]

[0050] In the formula, , , , , They are the mean square error (MSE) values ​​when completing the node injection voltage, node injection active power, node injection reactive power, branch active power, and branch current amplitude respectively; , , , , The samples are The true values ​​of node injection voltage, node injection active power, node injection reactive power, branch active power, and branch current amplitude are , , , , The samples are The node injection voltage, node injection active power, node injection reactive power, branch active power and branch current amplitude complement values;

[0051] When using multi-task learning to optimize topology identification, the cross entropy loss function is used As the loss function of the recognition task, the calculation expression is as follows:

[0052]

[0053] In the formula, The number of lines identified for a single training run, For Line The real category, line It takes 1 if it is in the closed state, otherwise it takes 0; It is the probability value of judging the line disconnection through the sigmoid function output; the overall multi-task learning loss function is composed of the weighted losses of each task, and the weights are controlled by trainable parameters. The expression is as follows:

[0054]

[0055] Where: is the overall loss function; , , , , , The loss functions for topology identification and measurement completion of node voltage, node injected active power, node injected reactive power, branch active power, and branch current are respectively trainable weight parameters. The trainable weight parameters are dynamically adjusted during the training process to achieve optimization of the overall model.

[0056] As a preferred solution, the steps of determining a multi-task learning target model, inputting real-time measurement data into the multi-task learning target model, and outputting a topology identification result include:

[0057] The Matpower simulation platform, a power system analysis toolbox in the matrix laboratory software Matlab, is used to perform power flow calculations by changing the topology structure, simulating and obtaining complete historical measurement data under different topology sets, and recording the corresponding topologies to jointly construct a training data set.

[0058] Construct network models for measurement completion and topology identification in multi-task learning networks, and determine the hyperparameters of each network model;

[0059] The training cycle begins by first using the constructed training data set as input to the autoencoder-based measurement completion network model, and performing encoding dimension reduction and decoding reconstruction operations on different measurement types to obtain the missing reconstruction values ​​of the measurement;

[0060] Based on the completed measurement data, the topology structure of the joint distribution network is modeled as graph data that can be learned by the graph neural network. The data is input into the pre-built topology identifier, the attention coefficient between the source node and the target node is calculated, and the source node and target node information are weighted. The mean of the node features under multi-head attention is taken as the output, and the line status is output through the fully connected neural network layer.

[0061] The optimization objectives of both measurement completion and topology identification are considered simultaneously to form the total loss function of the multi-task learning model. The adaptive moment estimation Adma optimization function is used to back-propagate and update the network. The model is evaluated to optimize the parameters.

[0062] Repeat the above steps until the multi-task learning model converges, save the optimal model parameters, and obtain the multi-task learning target model;

[0063] Real-time limited measurement data is input into the multi-task learning target model to output complete measurement data and topology identification results.

[0064] In a second aspect, an active distribution network topology identification system based on multi-task graph learning is provided, comprising:

[0065] A completion module is used to use a plurality of pre-built autoencoders to respectively complete the missing of different measurement features in the distribution network, and obtain the distribution network topology after the missing measurement features are completed;

[0066] A graph representation module is used to represent the distribution network topology after missing measurement features are completed in the form of a graph to obtain data in the form of a graph;

[0067] The feature matrix concatenation module is used to train the pre-built topology identifier using data in the form of a graph, and to aggregate messages from the target node to the source node according to the multi-head attention coefficient of the graph. The source node is the node currently being calculated, and each target node is the other node in the graph. The node association matrix is ​​used to calculate the feature vector difference of the head and tail nodes after the update, and the features of the head and tail nodes are concatenated to form the head and tail node feature vectors, and the two-dimensional features of the line itself are obtained. The feature vector difference of the head and tail nodes after the update is concatenated with the head and tail node feature vectors and the two-dimensional features of the line itself to form the final feature matrix of the edge of the entire graph.

[0068] The topology identification module is used to obtain the characteristics of each line according to the final feature matrix of the whole graph edge, input the characteristics of each line into the fully connected neural network layer, and map the characteristics into the judgment value of each line state. By setting the threshold of the judgment value for comparison, the disconnection state of the line is determined to achieve topology identification;

[0069] The multi-task learning optimization module is used to use multi-task learning to supplement the missing measurement features in the distribution network and the joint learning optimization of the topology identification process, determine the multi-task learning target model, input real-time measurement data into the multi-task learning target model, and output the topology identification results.

[0070] In a third aspect, a computer-readable storage medium is provided, wherein the computer-readable storage medium stores at least one instruction, and when the at least one instruction is executed by a processor, the active distribution network topology identification method based on multi-task graph learning is implemented.

[0071] Compared with the prior art, the first aspect of the present invention has at least the following beneficial effects:

[0072] The active distribution network topology identification method based on multi-task graph learning of the present invention takes into account the problem that the original measurement data may be missing. Before topology identification, an autoencoder network with an attention mechanism is first used to independently complete the limited measurement data available in the distribution network to ensure that each feature can be effectively reconstructed to obtain complete data for topology identification and restore the topological relationship information implicit in the measurement data, providing a rich and fine-grained embedded representation for downstream topology identification tasks, and improving the identification accuracy compared to a single measurement type. The present invention adds an improved graph algorithm using an attention mechanism to update and aggregate key measurement features, establishes a topology identification model based on binary classification of switch states, maps features to judgment values ​​for the operating state of each line / switch after filtering out irrelevant feature interference, and determines the disconnection state of the line by setting a threshold to achieve topology identification, thereby improving the model's reasoning ability for unknown topologies outside the generalization area. The present invention takes into account that the accuracy of topology identification is greatly affected by the integrity of the measurement data, and in order to ensure the consistency of the completed measurement data and the topology, accurate topology identification is also an inevitable problem in the measurement completion task. A two-stage multi-task training paradigm integrating measurement data completion and topology identification is proposed to form an end-to-end solution to simplify the model architecture and realize the robust identification of topology under limited measurement. In the topology identification task, compared with some deep learning, deep graph learning and machine learning algorithms, the present invention benefits from the forward iterative optimization between the two tasks. Even in the scenario with a high proportion of measurement missing, it can still maintain a high recognition accuracy and adapt to the input of different types of measurement data. At the same time, it also has strong generalization and noise resistance. In the task of measurement data completion, compared with the single-task measurement data completion without topology identification guidance, the multi-task joint learning proposed by the method of the present invention benefits from the auxiliary topology information guidance of the topology identification task. Even in the case of a high proportion of measurement data missing, it can restore the node and line measurements that conform to the real physical laws with high accuracy. Therefore, when faced with sparse and heterogeneous measurements in the distribution network, it can not only achieve efficient and robust identification of the topology, but also ensure the consistency of the completed measurement data and the topology to support the realization of online transparency of the situation.

[0073] It can be understood that the beneficial effects of the second to third aspects mentioned above can be found in the relevant description of the first aspect mentioned above, and will not be repeated here. BRIEF DESCRIPTION OF THE DRAWINGS

[0074] In order to more clearly illustrate the technical solutions in the embodiments of the present application, 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 application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0075] Figure 1 Architecture diagram of the active distribution network topology identification method based on multi-task graph learning in an embodiment of the present invention;

[0076] Figure 2 Schematic diagram of the topology identification workflow based on multi-task graph learning in an embodiment of the present invention. DETAILED DESCRIPTION

[0077] In the following description, specific details such as specific system structures, technologies, etc. are provided for the purpose of illustration rather than limitation, so as to provide a thorough understanding of the embodiments of the present application. However, it should be clear to those skilled in the art that the present application may also be implemented in other embodiments without these specific details. In other cases, detailed descriptions of well-known systems, devices, circuits, and methods are omitted to prevent unnecessary details from obstructing the description of the present application.

[0078] In the current field of distribution network topology identification, the distribution network topology identification method based on power flow matching is a method for identifying and verifying the structure of the distribution network, which relies on the simulation and analysis of the power system power flow (i.e., the flow of power in the network). The distribution network is simulated using a power flow calculation algorithm to calculate the voltage amplitude and phase angle of each node, as well as the power flow direction and size of each line; the actual measured power flow data is compared with the simulated power flow data, and the network parameters are adjusted through optimization algorithms (such as least squares method, genetic algorithm, etc.) to make the calculated power flow match the actual power flow as much as possible, and the topology with the highest matching degree is selected as the identification result. However, this method relies on the comprehensiveness of the historical topology library, the accuracy of state estimation cannot be guaranteed, and there may be a situation where the similarity of the same set of measured data is consistent with different topologies. In addition, the existing distribution network topology identification methods also use the electrical relationship and interdependence between nodes in the distribution network to identify the network structure, which usually does not rely on knowing the specific structure of the network in advance, but infers the topology through measurement data. This method determines the interdependence between nodes by analyzing the electrical relationship between nodes, such as voltage phase angle difference, power flow direction, etc. Then, it uses the node correlation information to infer the topology of the distribution network, identifies the node group through clustering algorithms, and verifies the inferred topology to ensure that it conforms to the actual power flow data. However, this method requires full node measurement data, and the actual branch nodes are usually not measured. This method is suitable for local branch line and phase identification, but it is difficult to identify the specific power supply path.

[0079] In view of the shortcomings of existing distribution network topology identification methods, in order to solve the challenge of topology identification caused by missing measurement data in distribution systems, an embodiment of the present invention provides an active distribution network topology identification method based on multi-task graph learning, thereby overcoming the limitations of traditional methods. The method of the embodiment of the present invention follows the two-stage multi-task training paradigm of measurement completion and topology identification. By comprehensively considering the optimization objectives of the two-stage tasks of measurement data completion and topology identification, the model parameters are jointly optimized in an end-to-end multi-task joint training manner to realize information interaction between the two tasks, thereby achieving a synergistic improvement in missing completion and identification effects, and then reducing the dimension, reconstructing incomplete data and accurately identifying the topology structure of the distribution network under the condition of limited measurement data.

[0080] See also Figure 1 and Figure 2 , the active distribution network topology identification method based on multi-task graph learning in the embodiment of the present invention mainly includes two stages, namely Task 1: measurement completion and Task 2: topology identification.

[0081] Among them, the measurement completion stage compresses and reconstructs the incomplete measurement data of the distribution network through the autoencoder; the topology identification stage extracts the measurement features through graph representation learning and performs binary classification of the switch state to achieve topology identification.

[0082] The active distribution network topology identification method based on multi-task graph learning in an embodiment of the present invention comprises the following steps:

[0083] Use multiple pre-built autoencoders to complete the missing of different measurement features in the distribution network, and obtain the distribution network topology after the missing measurement features are completed;

[0084] The distribution network topology after missing measurement features are completed is represented as a graph to obtain data in the form of a graph;

[0085] The pre-built topology identifier is trained using graph data, and messages are aggregated from the target node to the source node according to the multi-head attention coefficient of the graph. The source node is the node currently being calculated, and each target node is the other node in the graph. The node association matrix is ​​used to calculate the difference in the feature vectors of the head and tail nodes after the update, and the features of the head and tail nodes are concatenated to form the head and tail node feature vectors, and the two-dimensional features of the line itself are obtained. The feature vector difference of the head and tail nodes after the update is concatenated with the head and tail node feature vectors and the two-dimensional features of the line itself to form the final feature matrix of the edge of the entire graph.

[0086] According to the final feature matrix of the whole graph edge, the features of each line are obtained, and the features of each line are input into the fully connected neural network layer. The features are mapped to the judgment value of each line state, and the threshold of the judgment value is set for comparison to determine the disconnection state of the line and realize topology identification.

[0087] Multi-task learning is used to supplement the missing measurement features in the distribution network and optimize the joint learning of the topology identification process. The multi-task learning target model is determined, the real-time measurement data is input into the multi-task learning target model, and the topology identification results are output.

[0088] In one possible implementation, an autoencoder is a neural network model that follows an encoder-decoder and is used to learn a compact representation of data, and is typically used for tasks such as data completion, denoising, and feature learning. The embodiment of the present invention constructs multiple autoencoders to focus on processing the missing completion of different measurement features in the distribution network, ensuring that each feature can be effectively reconstructed. This independent processing method can better retain the important information of each feature and provide a rich and fine-grained measurement representation for downstream topology identification tasks. Specifically, the steps of using a plurality of pre-constructed autoencoders to respectively complete the missing of different measurement features in the distribution network in the embodiment of the present invention include:

[0089] In the encoding stage, the encoder fills the missing values ​​in the input measurement data with 0, extracts and captures the key information of the input features through the fully connected neural network layer, and outputs the encoded compressed representation. The expression is as follows:

[0090]

[0091] In the formula, is the output vector of the encoder, which represents the dense representation of the input measurement data of the distribution network; is the number of hidden layers; , The encoder is Trainable weight parameters and biases of the layer; For the l- The output vector of the 1-layer encoder, is the selected parameterized ReLU activation function;

[0092] In the decoding stage, the decoder receives the compressed representation generated by the encoder and generates a completion measurement similar to the original input through a neural network layer that is reversed from the fully connected neural network layer of the encoder. The expression is as follows:

[0093]

[0094] In the formula, For decoder The output vector of the layer, i.e. the reconstructed missing values; , Decoder No. Trainable weight parameters and biases of the layer; For decoder The output vector of the layer.

[0095] In a possible implementation manner, when the distribution network topology after the missing measurement features are completed is represented as a graph in the embodiment of the present invention, the graph is in the form of: , where is the set of distribution network nodes, Indicates the number of electrical nodes in the distribution network, is the set of edges of the graph, consisting of overhead lines or cable lines, using Represents the adjacency matrix of the graph; distribution network topology after missing measurement features are completed, where node voltage amplitude, node active power, and node reactive power are modeled as node features in the graph. Each node represents a bus in the power system, and the node feature is a three-dimensional vector , the current amplitude and branch active power are used as edge features in the graph, each edge represents a line in the power system, and the edge feature is a two-dimensional vector ; In this way, key information such as voltage, active power, and reactive power are retained at the same time. The initial characteristic matrix of the graph is , where Is a node The initial attribute information vector, Is A vector composed of initial attributes; for a graph data that needs to classify edges, the true labels of the edge states are constructed as a label matrix, expressed as ,in, is the number of categories, which is composed of the unique hot encoding of the categories. Finally, the true labels of the edge states and the graph topology are jointly modeled as complete graph data that can be used for training the graph neural network GNN model.

[0096] In a possible implementation, the topology identifier pre-built is trained using graph data in the embodiment of the present invention, and in the step of aggregating messages from the target node to the source node according to the multi-head attention coefficient of the graph, the topology identifier is built based on the Graph Transformer algorithm, and the node features in the initial graph data are After the Transformer algorithm is used to aggregate, the updated attribute information of all nodes is obtained. ; The measured characteristics of a given node Here, the measured features can be instantiated as node voltage amplitude, node injected active power, and node injected reactive power. Corresponding to the initial node features in the initial graph data, the following formula is used to calculate the initial node from each target node: To the source node The multi-head attention coefficient corresponding to each edge of :

[0097]

[0098]

[0099]

[0100]

[0101] In the formula, for the Attention head, l Layer-by-layer training weight matrices and the bias vector The node i Features Encoded as a query vector , training weight matrix and the bias vector The target node j Features Encoding key vector ; and train the weight matrix and the bias vector Edge feature Encoded as a relation vector Then added to the key vector as additional information for each layer; For source node and the target node The normalized attention weight coefficient between The numerator is the source node and the target node The attention value between the source nodes is The sum of the attention values ​​between all nodes in ; is the exponentially scaled dot product function, d is the hidden size of each attention head, is the number of nodes, and the query vector to the source node is obtained by point multiplication and the target node information;

[0102] According to the obtained multi-head attention coefficient, the multi-head attention matrix is ​​used as the transfer matrix of message passing, from the target node To the source node For message aggregation, the calculation expression is as follows:

[0103]

[0104]

[0105] In the formula, the weight matrix and the bias vector The target node Features Encoded as a vector of values After that, the relationship vector containing edge feature information Superposition, and then through the attention weight coefficient After weighted selection and aggregation, the next layer of source nodes is obtained Attention characteristics ,in Represents a splicing operation, is the number of attention heads, For source node Number.

[0106] In addition, in the training of the graph neural network, considering the transition smoothness problem of the model, the gated residual connection is used between the layers of the topology identifier to improve the layer structure. The calculation expression is as follows:

[0107]

[0108]

[0109]

[0110] In the formula, through the weight matrix and the bias vector The source node exist l Layer feature information Encoded as reset gating vector After that, with the source node exist l +1 layer of attention features The difference is encoded to obtain the updated gating vector ,in Represents the weight matrix Respectively , The weighted product is performed with the difference between the two, and sigmoid is used as the activation function; finally, the improved source node is obtained through the weighted product New Features ,in Representation layer normalization, For source node In updating the gating vector The strength of the corresponding residual connection in Simultaneous use l Layer features and l +1 layer output, increasing the diversity of features between different layers;

[0111] Finally, in order to obtain more comprehensive information and improve the stability of the model, the nonlinear transformation of the output is removed and the average is used to obtain the final node representation for topology identification.

[0112] After aggregation through the three-layer graph neural network GNN layer, the attribute information of all nodes is updated. , since the purpose of topology identification is to predict the switch state of the line, it is necessary to assign features to edge features. The final feature matrix of the edge of the entire graph described in the embodiment of the present invention is ,in , is the feature vector difference of the head and tail nodes, the feature vector of the head and tail nodes and Concatenate to form the head and tail node feature vectors , and They are the original features of the line and the features after completion. The feature dimension of each line in the embodiment of the present invention is 11. In order to obtain the state value of the line, it is input into the fully connected neural network layer and mapped to the judgment value of each line state through the Sigmoid function. , by setting the threshold to determine the disconnection status of the line, the threshold is set to 0.5. If the output judgment value If it is greater than 0.5, it is considered as 1, otherwise it is considered as 0, and the output value is binarized. The identification result can be expressed as ,in It is a binary variable, 1 represents that the line switch is in the closed state, and 0 represents that the line switch is in the open state, ultimately realizing topology identification.

[0113] In one possible implementation, Multi-Task Learning (MTL) is a machine learning method that allows a model to handle multiple independent but related tasks simultaneously. In the MTL framework, by sharing parameters in the model and combining the optimization objectives of multiple tasks, MTL can improve the training efficiency and generalization ability of the model. In the field of topology identification, the identification accuracy is affected by the measurement integrity. The measurement completion task can solve the topology identification problem under limited measurement. At the same time, topology identification can feed back to the measurement completion structural information to generate measurement information that is more in line with physical laws, thereby promoting more accurate topology identification. Therefore, the two tasks of measurement completion and topology identification are strongly correlated, and the measurement feature information is applicable to both tasks. Moreover, the implementation of the two tasks is mutually beneficial, and the accuracy can be improved based on the information provided by each other. Therefore, the use of the MTL framework can improve the accuracy of the topology identification model under limited measurement conditions.

[0114] The loss function expression of the multi-task learning target model described in the embodiment of the present invention is as follows:

[0115]

[0116] In the formula, , Indicates the first and The loss function of each task; , Represents subtasks The first predicted value and the true value; , Represents subtasks No. i Predicted values ​​and true values; Represents a linear combination of the loss functions of each subtask; during training, if you want a subtask to have a more prominent learning performance, you can use the subtask as the main task of the MTL framework and set a larger weight for the loss function of the task. In the model proposed in the embodiment of the present invention, the topology identification task is used as the main task of the MTL framework, and the measurement completion task is used to assist network learning.

[0117] When multi-task learning is used to optimize the learning of missing measurement features in the distribution network, the loss function is established with the mean square error MSE between the completed value and the true value as the object, and the calculation expression is as follows:

[0118]

[0119] In the formula, is the mean square error (MSE) value of measurement completion; is the number of samples; For sample The true value of For sample The complement value of ;

[0120] Different autoencoders are used for different measurement types, and the loss functions of each measurement are instantiated as follows:

[0121]

[0122]

[0123]

[0124]

[0125]

[0126] In the formula, , , , , They are the mean square error (MSE) values ​​when completing the node injection voltage, node injection active power, node injection reactive power, branch active power, and branch current amplitude respectively; , , , , The samples are The true values ​​of node injection voltage, node injection active power, node injection reactive power, branch active power, and branch current amplitude are , , , , The samples are The node injection voltage, node injection active power, node injection reactive power, branch active power and branch current amplitude complement values;

[0127] Since the training goal of topology identification is to determine the status of each line in the power system, the essence of this task is a binary classification problem. Therefore, when the embodiment of the present invention uses multi-task learning to optimize topology identification,

[0128] Using cross entropy loss function As the loss function of the recognition task, the calculation expression is as follows:

[0129]

[0130] In the formula, The number of lines identified for a single training run, For Line The real category, line It takes 1 if it is in the closed state, otherwise it takes 0; It is the probability value of judging the line disconnection through the sigmoid function output; the overall multi-task learning loss function is composed of the weighted losses of each task, and the weights are controlled by trainable parameters. The expression is as follows:

[0131]

[0132] Where: is the overall loss function; , , , , , The loss functions for topology identification and measurement completion of node voltage, node injected active power, node injected reactive power, branch active power, and branch current are respectively trainable weight parameters. The trainable weight parameters are dynamically adjusted during the training process to achieve optimization of the overall model.

[0133] In a possible implementation manner, the steps of determining a multi-task learning target model, inputting real-time measurement data into the multi-task learning target model, and outputting a topology identification result in the embodiment of the present invention include:

[0134] Step 1: Use the Matpower simulation platform of the power system analysis toolbox in the matrix laboratory software Matlab to perform power flow calculations by changing the topology structure, simulate and obtain the complete historical measurement data under different topology sets, and record the corresponding topology to jointly construct a training data set;

[0135] Step 2: Construct the network model of measurement completion and topology identification in the multi-task learning network, and determine the hyperparameters of each network model;

[0136] Step 3: The training cycle begins. First, the constructed training data set is input into the autoencoder-based measurement completion network model. Different measurement types are encoded and reduced in dimension, and decoded and reconstructed to obtain the missing reconstruction values ​​of the measurement.

[0137] Step 4: Based on the completed measurement data, the joint distribution network topology structure is modeled as graph data that can be learned by the graph neural network, and is input into the pre-built topology identifier. The attention coefficient between the source node and the target node is calculated, and the source node and target node information are weighted and aggregated. The node feature mean under multi-head attention is taken as the output, and the line status is output through the fully connected neural network layer;

[0138] Step 5: Consider the optimization objectives of both measurement completion and topology identification to form the total loss function of the multi-task learning model. Use the adaptive moment estimation Adma optimization function to back-propagate and update the network. At the same time, evaluate the model to optimize the parameters.

[0139] Step 6: Repeat steps 3 to 5 above until the multi-task learning model converges, save the optimal model parameters, and obtain the multi-task learning target model;

[0140] Step 7: Input real-time limited measurement data into the multi-task learning target model obtained in step 6 to output complete measurement data and topology identification results.

[0141] The active distribution network topology identification method based on multi-task graph learning in the embodiment of the present invention follows the two-stage multi-task training paradigm of measurement completion-topology identification, and jointly optimizes the model with end-to-end multi-task joint training, which can be used to reduce the dimension, reconstruct incomplete data and accurately identify the topology structure of the distribution network under the condition of limited measurement data. Compared with some deep learning, deep graph learning and machine learning algorithms, the method of the present invention benefits from the forward iterative optimization between the two tasks. Even in the scenario with a high proportion of missing measurement data, it can still maintain a high recognition accuracy and adapt to the input of different types of measurement data. It also has strong generalization and noise resistance. In the measurement data completion task, compared with the single-task measurement completion without topology identification guidance, the multi-task joint learning proposed by the method of the present invention benefits from the auxiliary topology information guidance of the topology identification task. Even in the case of a high proportion of missing measurement data, it can restore the node and line measurement data that conforms to the real physical laws with high accuracy. Therefore, when faced with sparse and heterogeneous measurements in the distribution network, it can not only achieve efficient and robust identification of the topology, but also ensure the consistency of the completed measurement data and the topology to support the realization of online transparency of the situation.

[0142] Another embodiment of the present invention further provides an active distribution network topology identification system based on multi-task graph learning, comprising:

[0143] A completion module is used to use a plurality of pre-built autoencoders to respectively complete the missing of different measurement features in the distribution network, and obtain the distribution network topology after the missing measurement features are completed;

[0144] A graph representation module is used to represent the distribution network topology after missing measurement features are completed in the form of a graph to obtain data in the form of a graph;

[0145] The feature matrix concatenation module is used to train the pre-built topology identifier using data in the form of a graph, and to aggregate messages from the target node to the source node according to the multi-head attention coefficient of the graph. The source node is the node currently being calculated, and each target node is the other node in the graph. The node association matrix is ​​used to calculate the feature vector difference of the head and tail nodes after the update, and the features of the head and tail nodes are concatenated to form the head and tail node feature vectors, and the two-dimensional features of the line itself are obtained. The feature vector difference of the head and tail nodes after the update is concatenated with the head and tail node feature vectors and the two-dimensional features of the line itself to form the final feature matrix of the edge of the entire graph.

[0146] The topology identification module is used to obtain the characteristics of each line according to the final feature matrix of the whole graph edge, input the characteristics of each line into the fully connected neural network layer, and map the characteristics into the judgment value of each line state. By setting the threshold of the judgment value for comparison, the disconnection state of the line is determined to achieve topology identification;

[0147] The multi-task learning optimization module is used to use multi-task learning to supplement the missing measurement features in the distribution network and the joint learning optimization of the topology identification process, determine the multi-task learning target model, input real-time measurement data into the multi-task learning target model, and output the topology identification results.

[0148] Another embodiment of the present invention further proposes an electronic device, including a processor and a memory, wherein the processor is used to execute a computer program stored in the memory to implement the active distribution network topology identification method based on multi-task graph learning.

[0149] Another embodiment of the present invention further proposes a computer-readable storage medium, wherein the computer-readable storage medium stores at least one instruction, and when the at least one instruction is executed by a processor, the active distribution network topology identification method based on multi-task graph learning is implemented.

[0150] The computer program includes computer program code, which may be in source code form, object code form, executable file or some intermediate form. The computer-readable storage medium may include: any entity or device, medium, USB flash drive, mobile hard disk, disk, optical disk, computer memory, read-only memory, random access memory, electric carrier signal, telecommunication signal and software distribution medium, etc. that can carry the computer program code. It should be noted that the content contained in the computer-readable medium can be appropriately increased or decreased according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media do not include electric carrier signals and telecommunication signals. For ease of explanation, the above content only shows the part related to the embodiment of the present invention. For specific technical details not disclosed, please refer to the method part of the embodiment of the present invention. The computer-readable storage medium is non-temporary and can be stored in a storage device formed by various electronic devices, which can realize the execution process recorded in the method of the embodiment of the present invention.

[0151] It should be understood by those skilled in the art that the embodiments of the present invention may be provided as methods, systems or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment or an embodiment combining software and hardware aspects. Moreover, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0152] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems) and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the process in the flowchart. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0153] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0154] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process in the computer or other programmable device. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0155] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the relevant field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents. Any modification or equivalent replacement that does not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for topology identification of active distribution network based on multi-task graph learning, characterized in that: include: Use multiple pre-built autoencoders to complete the missing of different measurement features in the distribution network, and obtain the distribution network topology after the missing measurement features are completed; The distribution network topology after missing measurement features are completed is represented as a graph to obtain data in the form of a graph; The pre-built topology identifier is trained using graph data, and messages are aggregated from the target node to the source node according to the multi-head attention coefficient of the graph. The source node is the node currently being calculated, and each target node is the other nodes in the graph. The node association matrix is ​​used to calculate the difference in feature vectors of the head and tail nodes after the update, and the features of the head and tail nodes are concatenated to form the feature vectors of the head and tail nodes, and the two-dimensional features of the line itself are obtained. The difference in feature vectors of the head and tail nodes after the update is concatenated with the feature vectors of the head and tail nodes and the two-dimensional features of the line itself to form the final feature matrix of the edge of the entire graph. According to the final feature matrix of the whole graph edge, the features of each line are obtained, and the features of each line are input into the fully connected neural network layer. The features are mapped to the judgment value of each line state, and the threshold of the judgment value is set for comparison to determine the disconnection state of the line and realize topology identification. Multi-task learning is used to supplement the missing measurement features of different distribution networks and to optimize the joint learning of the topology identification process. The multi-task learning target model is determined, the real-time measurement data is input into the multi-task learning target model, and the topology identification results are output.

2. The active distribution network topology identification method based on multi-task graph learning according to claim 1 is characterized in that: The steps of using a plurality of pre-built autoencoders to respectively complete the missing of different measurement features in the distribution network include: In the encoding stage, the encoder fills the missing values ​​in the input measurement data with 0, extracts and captures the key information of the input features through the fully connected neural network layer, and outputs the encoded compressed representation. The expression is as follows: In the formula, for l The output vector of the layer encoder represents the dense representation of the input measurement data of the distribution network; , The encoder is Trainable weight parameters and biases of the layer; For the l- The output vector of the 1-layer encoder, is the selected parameterized ReLU activation function; In the decoding stage, the decoder receives the compressed representation generated by the encoder and generates a completion measurement similar to the original input through a neural network layer that is reversed from the fully connected neural network layer of the encoder. The expression is as follows: In the formula, For decoder The output vector of the layer, i.e. the reconstructed missing values; , Decoder No. Trainable weight parameters and biases of the layer; For decoder The output vector of the layer.

3. The active distribution network topology identification method based on multi-task graph learning according to claim 1 is characterized in that: When the distribution network topology after the missing measurement features are completed is represented in the form of a graph, the form of the graph is , where is the set of distribution network nodes, Indicates the number of electrical nodes in the distribution network, is the set of edges of the graph, consisting of overhead lines or cable lines, using Represents the adjacency matrix of the graph; distribution network topology after missing measurement features are completed, where node voltage amplitude, node active power, and node reactive power are modeled as node features in the graph. Each node represents a bus in the power system, and the node feature is a three-dimensional vector , the current amplitude and branch active power are used as edge features in the graph, each edge represents a line in the power system, and the edge feature is a two-dimensional vector ; The initial feature matrix of the graph is , where Is a node The initial attribute information vector, Is A vector composed of initial attributes; for a graph data that needs to classify edges, the true labels of the edge states are constructed as a label matrix, expressed as ,in, is the number of categories for classification, consisting of the one-hot encoding of the categories.

4. The active distribution network topology identification method based on multi-task graph learning according to claim 1 is characterized in that: In the step of training the pre-built topology identifier using the graph data and aggregating messages from the target node to the source node according to the multi-head attention coefficient of the graph, the topology identifier is built based on the Graph Transformer algorithm, and the node features in the initial graph data are After algorithm aggregation, the updated attribute information of all nodes is obtained ; The measured characteristics of a given node , corresponding to the initial node features in the initial graph data, is calculated from each target node as follows To the source node The multi-head attention coefficient corresponding to each edge of : In the formula, for the Attention head, l Layer-by-layer training weight matrices and the bias vector The node i Features Encoded as a query vector , training weight matrix and the bias vector The target node j Features Encoding key vector ; and train the weight matrix and the bias vector Edge Features Encoded as a relation vector Then added to the key vector as additional information for each layer; For source node and the target node The normalized attention weight coefficient between The numerator is the source node and the target node The attention value between the source nodes is The sum of the attention values ​​between all nodes in ; is the exponentially scaled dot product function, d is the hidden size of each attention head, is the number of nodes, and the query vector to the source node is obtained by point multiplication and the target node information; According to the obtained multi-head attention coefficient, the multi-head attention matrix is ​​used as the transfer matrix of message passing, from the target node To the source node For message aggregation, the calculation expression is as follows: In the formula, the weight matrix and the bias vector The target node Features Encoded as a vector of values After that, the relationship vector containing edge feature information Superposition, and then through the attention weight coefficient After weighted selection and aggregation, the next layer of source nodes is obtained Attention characteristics ,in Represents a splicing operation, is the number of attention heads, For source node Number.

5. The active distribution network topology identification method based on multi-task graph learning according to claim 4 is characterized in that: It also includes using gated residual connections between the layers of the topology identifier to improve the layer structure. The calculation expression is as follows: In the formula, through the weight matrix and the bias vector The source node exist l Layer feature information Encoded as reset gating vector After that, with the source node exist l +1 layer of attention features The difference is encoded to obtain the updated gating vector ,in Represents the weight matrix Respectively , The weighted product is performed with the difference between the two, and sigmoid is used as the activation function; finally, the improved source node is obtained through the weighted product New Features ,in Representation layer normalization, The source node In updating the gating vector The strength of the corresponding residual connection in Simultaneous use l Layer features and l +1 layer output, increasing the diversity of features between different layers; It also includes removing nonlinear transformation from the output and taking an average approach to obtain a final node representation for topology identification.

6. The active distribution network topology identification method based on multi-task graph learning according to claim 1 is characterized in that: The final feature matrix of the entire graph edge is ,in , is the feature vector difference of the head and tail nodes, the feature vector of the head and tail nodes and Concatenate to form the head and tail node feature vectors , and They are the original features of the line and the features after completion; The Sigmoid function is used to map the features into the judgment value of each line state. , the judgment value The set threshold is compared, and if it is greater than the threshold, it is considered to be 1, otherwise it is considered to be 0, where 1 represents that the circuit breaker is in a closed state, and 0 represents that the circuit breaker is in an open state.

7. The active distribution network topology identification method based on multi-task graph learning according to claim 1 is characterized in that: The loss function expression of the multi-task learning target model is as follows: In the formula, , Indicates the first and The loss function of each task; , Represents subtasks The first predicted value and the true value; , Represents subtasks No. i Predicted values ​​and true values; Represents the linear combination of the loss functions of each subtask; When multi-task learning is used to optimize the learning of missing measurement features in the distribution network, the loss function is established with the mean square error MSE between the completed value and the true value as the object, and the calculation expression is as follows: In the formula, is the mean square error (MSE) value of measurement completion; is the number of samples; For sample The true value of For sample The complement value of ; Different autoencoders are used for different measurement types, and the loss functions of each measurement are instantiated as follows: In the formula, , , , , They are the mean square error (MSE) values ​​when completing the node injection voltage, node injection active power, node injection reactive power, branch active power, and branch current amplitude respectively; , , , , The samples are The true values ​​of node injection voltage, node injection active power, node injection reactive power, branch active power, and branch current amplitude are , , , , The samples are The node injection voltage, node injection active power, node injection reactive power, branch active power and branch current amplitude complement values; When using multi-task learning to optimize topology identification, the cross entropy loss function is used As the loss function of the recognition task, the calculation expression is as follows: In the formula, The number of lines identified for a single training run, For Line The real category, line It takes 1 if it is in the closed state, otherwise it takes 0; It is the probability value of judging the line disconnection through the sigmoid function output; the overall multi-task learning loss function is composed of the weighted losses of each task, and the weights are controlled by trainable parameters. The expression is as follows: Where: is the overall loss function; , , , , , The loss functions for topology identification and measurement completion of node voltage, node injected active power, node injected reactive power, branch active power and branch current are respectively trainable weight parameters. The trainable weight parameters are dynamically adjusted during the training process to achieve optimization of the overall model.

8. The active distribution network topology identification method based on multi-task graph learning according to claim 1 is characterized in that: The steps of determining a multi-task learning target model, inputting real-time measurement data into the multi-task learning target model, and outputting a topology identification result include: The Matpower simulation platform, a power system analysis toolbox in the matrix laboratory software Matlab, is used to perform power flow calculations by changing the topology structure, simulating and obtaining complete historical measurement data under different topology sets, and recording the corresponding topologies to jointly construct a training data set. Construct network models for measurement completion and topology identification in multi-task learning networks, and determine the hyperparameters of each network model; The training cycle begins by first using the constructed training data set as input to the autoencoder-based measurement completion network model, and performing encoding dimension reduction and decoding reconstruction operations on different measurement types to obtain the missing reconstruction values ​​of the measurement; Based on the completed measurement data, the topology structure of the joint distribution network is modeled as graph data that can be learned by the graph neural network. The data is input into the pre-built topology identifier, the attention coefficient between the source node and the target node is calculated, and the source node and target node information are weighted. The mean of the node features under multi-head attention is taken as the output, and the line status is output through the fully connected neural network layer. The optimization objectives of both measurement completion and topology identification are considered simultaneously to form the total loss function of the multi-task learning model. The adaptive moment estimation Adma optimization function is used to back-propagate and update the network. The model is evaluated to optimize the parameters. Repeat the above steps until the multi-task learning model converges, save the optimal model parameters, and obtain the multi-task learning target model; Real-time limited measurement data is input into the multi-task learning target model to output complete measurement data and topology identification results.

9. An active distribution network topology identification system based on multi-task graph learning, characterized in that: include: A completion module is used to use a plurality of pre-built autoencoders to respectively complete the missing of different measurement features in the distribution network, and obtain the distribution network topology after the missing measurement features are completed; A graph representation module is used to represent the distribution network topology after missing measurement features are completed in the form of a graph to obtain data in the form of a graph; The feature matrix concatenation module is used to train the pre-built topology identifier using graph data, and to aggregate messages from the target node to the source node according to the multi-head attention coefficient of the graph. The source node is the node currently being calculated, and each target node is the other nodes in the graph. The node association matrix is ​​used to calculate the difference in feature vectors between the head and tail nodes after the update, and the features of the head and tail nodes are concatenated to form the feature vectors of the head and tail nodes, and the two-dimensional features of the line itself are obtained. The difference in feature vectors between the head and tail nodes after the update is concatenated with the feature vectors of the head and tail nodes and the two-dimensional features of the line itself to form the final feature matrix of the edge of the entire graph. The topology identification module is used to obtain the characteristics of each line according to the final feature matrix of the whole graph edge, input the characteristics of each line into the fully connected neural network layer, and map the characteristics into the judgment value of each line state. By setting the threshold of the judgment value for comparison, the disconnection state of the line is determined to achieve topology identification; The multi-task learning optimization module is used to use multi-task learning to supplement the missing measurement features in the distribution network and the joint learning optimization of the topology identification process, determine the multi-task learning target model, input real-time measurement data into the multi-task learning target model, and output the topology identification results.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores at least one instruction, and when the at least one instruction is executed by the processor, it implements the active distribution network topology identification method based on multi-task graph learning as described in any one of claims 1 to 8.

Images