Method for Fault Diagnosis and Location of Distribution Network Based on Super-Resolution and Graph Neural Network

Abstract

The present invention relates to the technology of distribution network fault diagnosis and location, aiming to provide a distribution network fault diagnosis and location method based on super-resolution and graph neural network. It includes: collecting the fault feature data of key nodes in the distribution network, using a super-resolution model based on graph convolutional network to reconstruct the data of all nodes in the entire distribution network, and obtaining the estimated value of the full-node feature data of the distribution network under fault conditions; using the data estimate and continuing to use a fault diagnosis and location model based on graph attention network to obtain the fault type and the location of the fault node in the distribution network. The present invention applies graph neural network to the field of distribution network fault diagnosis and location, improves the effectiveness and accuracy of distribution network fault diagnosis and location, and can improve the operation stability and reliability of the distribution network; the present invention uses super-resolution technology for the feature reconstruction of the distribution network before fault diagnosis and location, effectively reduces the deployment of distribution network intelligent acquisition devices, and greatly reduces the construction cost of the distribution network.

Description

Technical Field

[0001] The present invention relates to the technology of distribution network fault diagnosis and location, and specifically to a method for distribution network fault diagnosis and location based on super-resolution and graph neural network. Background Art

[0002] The reliability of the distribution network is the key to ensuring the safe and stable power supply to users. To improve its reliability, system operators must promptly handle faults. Therefore, it is particularly important to accurately and quickly locate and clear faults immediately after they occur. And to facilitate the operator to correctly clear the faults, it is also crucial to accurately classify the faults.

[0003] In recent years, more and more distributed power sources have been connected to the distribution network, posing challenges to traditional fault diagnosis and location methods. With the rise of artificial intelligence technology, more and more artificial intelligence-based methods have been proposed for distribution network fault diagnosis and location. These methods weaken the influence of load changes and avoid the injection of high-frequency signals, achieving certain results in this field. However, most of these new artificial intelligence-based methods are based on wide-area measurement assisted by intelligent devices, and the economic cost of using them for the deployment of intelligent acquisition devices is relatively high. Moreover, due to the relatively simple models, the performance of many methods based on traditional machine learning models is not good enough. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to overcome the deficiencies in the prior art and provide a method for distribution network fault diagnosis and location based on super-resolution and graph neural network.

[0005] To solve the technical problem, the solution of the present invention is:

[0006] Provide a method for distribution network fault diagnosis and location based on super-resolution and graph neural network, including the following steps:

[0007] (1) Collect fault feature data of key nodes in the distribution network based on a micro-phasor measurement unit (μPMU); the key nodes refer to the nodes directly connected to the external power grid and distributed power sources, or the nodes directly connected to at least three other nodes;

[0008] (2) Based on the collected fault feature data, use a super-resolution model based on a graph convolutional network (GCN) to reconstruct the data of all nodes in the entire distribution network to obtain the estimated value of the full-node feature data of the distribution network under the fault state;

[0009] (3) Based on the estimated value of the full-node feature data of the distribution network under the fault state, use a fault diagnosis and location model based on a graph attention network (GAT) to obtain the fault type and the location of the fault node in the distribution network.

[0010] As a preferred solution of the present invention, step (2) specifically includes:

[0011] (2.1) Obtain the characteristic data of key nodes and all nodes under the fault state of the distribution network as the input and output samples for the training of the super-resolution model;

[0012] (2.2) Build a super-resolution model based on the graph convolutional network;

[0013] (2.3) Use the gradient descent algorithm to iteratively train the model until the loss converges;

[0014] (2.4) Input the key node information under the fault state of the distribution network into the trained model to obtain the estimated value of the characteristic data of all nodes in the distribution network.

[0015] As a preferred solution of the present invention, step (3) specifically includes:

[0016] (3.1) Take the fault type and location label of the distribution network and the estimated value of the characteristic data of all nodes in the distribution network under the fault state obtained in step (2) as the output and input samples for model training;

[0017] (3.2) Build a fault diagnosis and location model based on the graph attention network to realize the two functions of fault diagnosis and fault node location;

[0018] (3.3) Use the input and output samples to train the model for the two parts of fault diagnosis and fault node location under different fault types respectively;

[0019] (3.4) Input the estimated value of the characteristic data of all nodes in the distribution network under the fault state into the trained fault diagnosis and location model, and obtain the fault type of the distribution network after fault diagnosis;

[0020] (3.5) According to the obtained fault type, input the estimated value of the characteristic data of all nodes in the distribution network under the fault state into the fault diagnosis and location model again for corresponding fault node location to obtain the fault node location of the distribution network.

[0021] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0022] 1. The present invention applies the graph neural network to the field of distribution network fault diagnosis and location. Since the distribution network topology is a graph topology structure, the method based on the graph neural network is very suitable for this field. The present invention improves the effectiveness and accuracy of distribution network fault diagnosis and location, and can improve the operation stability and reliability of the distribution network.

[0023] 2. The present invention introduces super-resolution technology into the field of distribution network fault diagnosis and location, and is used for reconstructing the characteristics of the distribution network before fault diagnosis and location. This makes the present invention different from most existing artificial intelligence methods that require wide-area measurement of all nodes in the distribution network. Only intelligent acquisition devices need to be installed at some key nodes, effectively reducing the deployment of intelligent acquisition devices in the distribution network and significantly reducing the construction cost of the distribution network. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 is a simple flow chart implemented by the present invention;

[0025] Figure 2 is the topology diagram of the IEEE 37-node system;

[0026] Figure 3 is the structural diagram of the super-resolution model based on the graph convolutional network (GCN);

[0027] Figure 4 is the structural diagram of the unified model for fault diagnosis and location based on the graph attention network (GAT);

[0028] Figure 5 is the flow chart of fault diagnosis and location. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0029] First of all, it should be noted that the present invention relates to big data and machine learning technologies, and is an application of computer technology in the field of industrial control technology. In the implementation process of the present invention, the application of multiple software function modules will be involved. The applicant believes that after carefully reading the application documents and accurately understanding the implementation principle and invention purpose of the present invention, and in combination with the existing well-known technologies, those skilled in the art can fully implement the present invention by using their mastered software programming skills. The aforementioned software function modules include but are not limited to: the super-resolution model based on the graph convolutional network, the fault diagnosis and location model based on the graph attention network, etc. All those mentioned in the application documents of the present invention belong to this category, and the applicant will not list them one by one.

[0030] Those skilled in the art know that in addition to implementing a part of the system provided by the present invention and its various devices, modules, and units in the form of pure computer-readable program code, the method steps can be logically programmed to make the system provided by the present invention and its various devices, modules, and units be implemented in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers, etc. to achieve the same function. Therefore, the system provided by the present invention and its various devices, modules, and units can be regarded as a hardware component, and the devices, modules, and units included therein for implementing various functions can also be regarded as the structure within the hardware component; the devices, modules, and units for implementing various functions can also be regarded as both software modules for implementing the method and the structure within the hardware component.

[0031] It should also be noted that, in this text, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the term "comprising", "including" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, article or device comprising a series of elements not only includes those elements, but also includes other elements not expressly listed, or also includes elements inherent to such process, method, article or device. Without further limitation, an element defined by the statement "comprising one..." does not exclude the presence of additional identical elements in the process, method, article or device comprising the said element.

[0032] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments. The example implementation scenario is the IEEE 37-node system, and the system fault data are all generated by simulation using OpenDSS software.

[0033] The implementation process of the present invention is as follows: First, according to the characteristic data of the key nodes of the distribution network, the super-resolution model based on the graph convolutional network (GCN) reconstructs the full-node characteristic data; based on the estimated value of the full-node characteristic data of the reconstructed distribution network, the fault diagnosis and location model first judges the fault type, and then inputs the estimated value of the full-node characteristic data into the fault diagnosis and location model according to the fault type to obtain the corresponding fault node location. The fault diagnosis and location model used is a unified model based on the graph attention network (GAT).

[0034] The specific steps of the embodiment are as follows:

[0035] Step 1: Build an IEEE 37-node system model in OpenDSS software, and its topological structure is as Figure 2 shown. According to the definition of key nodes, the key node data are measured from μPMUs installed on nodes 702, 703, 704, 705, 707, 708, 709, 710, 711, 720, 734, 744, 799. Based on OpenDSS software, fault simulations are performed on all nodes in the IEEE 37-node system. The default fault resistance is 10 ohms, and the load level is randomly selected between 0.3 and 1. During the fault, voltage, current phasors and power are measured to obtain the training dataset and test dataset for feature reconstruction and fault diagnosis and location.

[0036] In this embodiment, fault simulation is carried out by building a model in software. In the actual operation environment of the distribution network, a micro - phasor measurement unit (μPMU) is used to collect fault feature data of key nodes in the distribution network. The micro - phasor measurement unit belongs to the prior art, and no special requirements are made in this invention; the specific data acquisition method is a skill mastered by those skilled in the art, and this invention will not elaborate further.

[0037] It is defined that the fault types include single - line - to - ground fault (SLG), line - to - line fault (LL), and line - to - line - to - ground fault (LLG). 50 data samples are generated for each fault type and the non - fault type on each node. Therefore, the entire data set contains 7400 data samples, which are divided into an 80% training set and a 20% test set.

[0038] Step 2: Based on the training data set, build and train a super - resolution model based on the graph convolutional network (GCN) and a fault diagnosis and location model based on the graph attention network (GAT) respectively.

[0039] Step 2.1: Build and train a super - resolution model based on the graph convolutional network (GCN), specifically as follows:

[0040] (1) Build a super - resolution model based on the graph convolutional network (GCN), the structure is as Figure 3 shown, and the specific method is as follows:

[0041] Use the super - resolution model based on the graph convolutional network (GCN) to reconstruct the distribution network features. Its input is the input matrix X containing the feature information of key nodes in the distribution network and the distribution network topology adjacency matrix A, expressed as:

[0042] Input=(X, A)

[0043] In the formula, X is an N×F - dimensional matrix composed of the features x i of each node i. N is the number of nodes, and F is the number of features. Among them, the features x i of key nodes are filled with real measurement data, and the features x i of non - key nodes are filled with the features of the key node closest to them;

[0044] The output label of the super - resolution model is an N×F - dimensional matrix Y, which is composed of the real feature data measurements y i of each node i;

[0045] (2) The super - resolution model consists of two graph convolutional layers and one fully - connected layer. The expression of each graph convolutional layer is as follows:

[0046]

[0047] In the formula, H(l+1) and H (l) represent the outputs of the (l + 1)-th and l-th graph convolutional layers; is the weight matrix, where F h is the output dimension of the graph convolutional layer; A is the adjacency matrix of the distribution network topology; A = A + I, where I is the identity matrix; D is the degree matrix of A; σ(.) is the activation function;

[0048] Define Then the output of the fully connected layer is expressed as:

[0049] Z i = W f σ(Aσ(AX i W (1) )W (2) ) + b f

[0050] In the formula, and are the weight matrix and bias matrix of the fully connected layer respectively; X i is the feature matrix of the i-th node; W (1) and W (2) are the weight matrices of the first and second graph convolutional layers respectively;

[0051] The above activation function σ(.) all adopts the LeakyReLU function, and its expression is as follows:

[0052] σ(x) = LeakyReLU(x, β) = max(0, x) + β × min(0, x)

[0053] In the formula, x is the input of the function;

[0054] The super-resolution model will be trained in a supervised manner, and its loss function is composed of the mean square error (MSE) and the Kullback-Leibler divergence loss (KLDivLoss), and the expression is as follows:

[0055]

[0056] In the formula, L represents the loss value, y i is the measurement of the true feature data of each node i; z i is the output feature of the model for each node i; N is the number of nodes.

[0057] (3) Use the key node feature data and all-node feature data under the fault state of the distribution network in the training dataset as the input and output samples for model training;

[0058] (4) Use the gradient descent algorithm to iteratively train the super-resolution model until the loss converges.

[0059] Step 2.2: Build and train a fault diagnosis and location model based on the Graph Attention Network (GAT) as follows:

[0060] (1) Build a unified fault diagnosis and location model based on the Graph Attention Network (GAT) to achieve the two functions of fault diagnosis and fault node location. The structure of the model is as Figure 4 shown. The specific building method is as follows:

[0061] Use a model based on the Graph Attention Network (GAT) for fault diagnosis and location. Its input is the input matrix X composed of the estimated values of the full-node feature data under the fault state of the distribution network and the topological adjacency matrix A of the distribution network, expressed as:

[0062] Input = (X, A)

[0063] where X is an F×N-dimensional matrix composed of the feature estimates x i of each node i, N is the number of nodes, and F is the number of features;

[0064] The output labels of the model include two parts, namely the fault type label Yc and the fault location label Yl;

[0065] (2) The fault diagnosis and location model is a unified model that can achieve the unification of the two functions of fault diagnosis and fault node location. Each functional unit consists of two graph attention layers and a fully connected layer, only the output dimension of the model is different. The expression of each graph attention layer is as follows:

[0066]

[0067] where h′ represents the output of the graph attention layer of all nodes, represents the output of the graph attention layer of the i-th node, and || represents matrix concatenation; is the normalized attention coefficient calculated by the k-th attention mechanism; is the corresponding weight matrix; is the set of all neighbor nodes of node i; K is the number of heads in the attention mechanism; σ(.) is the activation function LeakyReLU; N is the number of nodes;

[0068] The above attention mechanism is a single-layer feedforward neural network parameterized by the weight vector and its expression is:

[0069]

[0070] where is the weight matrix; soft max(.) is the normalization exponential function; exp(.) is the exponential function with base e; represents the output of the graph attention layer of the i-th node; represents the weight vector of the attention mechanism network;

[0071] The output of the fully connected layer (i.e., the output of the fault diagnosis and location model) is calculated by the following formula:

[0072]

[0073]

[0074]

[0075] where, and are the feature outputs of the i-th node of the first and second graph attention layers respectively; σ(.) is the activation function; K1 and K2 are the numbers of heads in the attention mechanism in the first and second graph attention layers respectively; is the feature input of the j-th node; and are the normalized attention coefficients of the first and second graph attention layers respectively; and are the weight matrices of the first and second graph attention layers respectively, where F h is the output dimension of the graph attention layer; Z is the output of the fully connected layer; N is the number of nodes; and are the weight matrix and bias matrix of the fully connected layer respectively, where N o is the output dimension of the fully connected layer;

[0076] (3) The fault diagnosis and location model will be trained in a supervised manner, and its loss function uses cross-entropy loss, and the expression is as follows:

[0077]

[0078] where, onehot(.) represents the one-hot vectorization function, L is the loss value, y i is the fault type or location label, z i is the output of the fault diagnosis and location model.

[0079] (4) Use the estimated values of all node feature data and the corresponding distribution network fault type labels in the training dataset as the input and output samples for model training, and use the gradient descent algorithm to iteratively train the model until the loss converges;

[0080] (5) Use the estimated full-node feature data and the corresponding distribution network fault location labels in the training dataset as the input-output samples for model training. Train the model separately for fault location under different fault types, and use the gradient descent algorithm for iterative training until the loss converges.

[0081] Step 3: Obtain the key node feature data under the fault state of the distribution network in the test dataset, and input it into the super-resolution model trained in Step 2.1 to obtain the estimated value of the full-node feature data under the reconstructed distribution network fault state.

[0082] Step 4: According to Figure 5 the shown fault diagnosis and location process, input the estimated value of the full-node feature data under the distribution network fault state obtained into the fault diagnosis and location model trained in Step 2.2 to obtain the distribution network fault type. Then, according to the fault type, input the estimated value of the full-node feature data into the fault diagnosis and location model again for corresponding fault node location, and the accurate location of the distribution network fault node can be obtained.

Claims

1. A method for fault diagnosis and location of a distribution network based on super-resolution and graph neural network, characterized in that, It includes the following steps: (1) Collect fault feature data of key nodes in the distribution network based on a micro - phasor measurement unit; the key nodes refer to the nodes directly connected to the external power grid and distributed power sources, or the nodes directly connected to at least three other nodes; (2) Based on the collected fault feature data, use a super - resolution model based on a graph convolutional network to reconstruct all - node data of the entire distribution network, and obtain the estimated values of all - node feature data of the distribution network in the fault state; (3) Based on the estimated values of all - node feature data of the distribution network in the fault state, use a fault diagnosis and location model based on a graph attention network to obtain the fault type and the location of the fault node in the distribution network.

2. The method according to claim 1, wherein The specific steps of step (2) include: (2.1) Obtain the feature data of key nodes and all nodes in the fault state of the distribution network as the input - output samples for training the super - resolution model; (2.2) Build a super - resolution model based on a graph convolutional network; (2.3) Use the gradient descent algorithm to iteratively train the model until the loss converges; (2.4) Input the key - node information in the fault state of the distribution network into the trained model to obtain the estimated values of all - node feature data of the distribution network.

3. The method according to claim 2, wherein The specific steps of step (2.2) include: (2.2.1) Take the input matrix X containing the feature information of key nodes in the distribution network and the topological adjacency matrix A of the distribution network as the input of the super - resolution model, expressed as: Input=(X, A) Wherein, X is the feature x of each node i i to form an N×F dimensional matrix, where N is the number of nodes and F is the number of features. Among them, the feature x of the key node i is filled with real measurement data, and the feature x of the non-key node i is filled with the feature of the key node closest to it; The output label of the super-resolution model is an N×F-dimensional matrix Y, which is composed of the true feature data measurements y of each node i i ; (2.2.2) The super - resolution model consists of two graph convolutional layers and a fully - connected layer. The expression of each graph convolutional layer is as follows: where H (l+1) and H (l) represent the outputs of the (l + 1)-th and l-th graph convolutional layers; W (l) is the weight matrix and where F h is the output dimension of the graph convolutional layer; A is the adjacency matrix of the distribution network topology; A = A + I, where I is the identity matrix; D is the degree matrix of A; σ(.) is the activation function; Definition The output of the fully connected layer is represented as: Z i = W f σ(Aσ(AX i W (1) )W (2) ) + b f Where, W f and b f are the weight matrix and bias matrix of the fully connected layer, respectively, and X i is the feature matrix of the i-th node; W (1) and W (2) are the weight matrices of the first and second graph convolutional layers, respectively; The above activation function σ(.) all adopts the LeakyReLU function, and its expression is as follows: σ(x)=LeakyReLU(x,β)=max(0,x)+β×min(0,x) where x is the input of the function; (2.2.3) The super - resolution model will be trained in a supervised manner, and its loss function consists of mean - square error and Kullback - Leibler divergence loss, and the expression is as follows: Where L represents the loss value, and y i is the measured true feature data for each node i; z i is the model output feature for each node i; and N is the number of nodes.

4. The method according to claim 1, wherein The specific steps of step (3) include: (3.1) Take the fault type and location labels of the distribution network and the estimated values of all - node feature data of the distribution network in the fault state obtained in step (2) as the output and input samples for model training; (3.2) Build a fault diagnosis and location model based on a graph attention network to achieve the two functions of fault diagnosis and fault - node location; (3.3) Use the input - output samples to train the model for the two parts of fault diagnosis and fault - node location under different fault types respectively; (3.4) Input the estimated values of all - node feature data of the distribution network in the fault state into the trained fault diagnosis and location model, and obtain the fault type of the distribution network after fault diagnosis; (3.5) According to the obtained fault type, input the estimated values of all - node feature data of the distribution network in the fault state into the fault diagnosis and location model again for corresponding fault - node location to obtain the location of the fault node in the distribution network.

5. The method according to claim 4, wherein The specific steps of step (3.2) include: (3.2.1) The input matrix X composed of the estimated values of the full-node characteristic data under the fault state of the distribution network and the distribution network topology adjacency matrix A are used as the inputs of the fault diagnosis and location model, expressed as: Input=(X, A) where X is an F×N dimensional matrix composed of the feature estimates x of each node i i , N is the number of nodes, and F is the number of features; The output labels of the model include two parts, namely the fault type label Yc and the fault location label Yl; (3.2.2) The fault diagnosis and location model is a unified model that can achieve the two functions of fault diagnosis and fault node location. Each functional unit is composed of two graph attention layers and a fully connected layer, only the output dimension of the model is different. The expression of each graph attention layer is as follows: where h′ represents the output of the graph attention layer for all nodes, represents the output of the graph attention layer for the i-th node, and || represents matrix concatenation; is the normalized attention coefficient, calculated by the k-th attention mechanism; W k is the weight matrix, and is the set of all neighbor nodes of node i; K is the number of heads in the attention mechanism; σ(.) is the activation function LeakyReLU; N is the number of nodes; The above attention mechanism is a single-layer feedforward neural network parameterized by a weight vector and its expression is as follows: where W is the weight matrix, and softmax(.) is the normalized exponential function; exp(.) is the exponential function with base e; represents the output of the graph attention layer of the i-th node; represents the weight vector of the attention mechanism network; (3.2.3) The output of the fully connected layer is calculated by the following formula: wherein, and are the feature outputs of the i-th node of the first and second graph attention layers respectively; σ(.) is the activation function; K1 and K2 are the numbers of heads in the attention mechanism of the first and second graph attention layers respectively; is the feature input of the j-th node; and are the normalized attention coefficients of the first and second graph attention layers respectively; and are the weight matrices of the first and second graph attention layers respectively, and where F h is the output dimension of the graph attention layer; Z is the output of the fully connected layer; N is the number of nodes; W f and b f are the weight matrix and bias matrix of the fully connected layer respectively, and where N o is the output dimension of the fully connected layer; (3.2.4) The fault diagnosis and location model will be trained in a supervised manner, and its loss function uses cross-entropy loss, and the expression is as follows: where onehot(.) represents the one-hot vectorization function, L is the loss value, y i is the fault type or location label, and z i is the output of the fault diagnosis and location model.

Images