Novel power distribution network online accident analysis method and system based on graph convolutional neural network

Abstract

The invention discloses a novel power distribution network online accident analysis method and system based on a graph convolutional neural network. The novel power distribution network on-line accident analysis method comprises the following steps: forming a node characteristic data matrix by the node voltage, current and power of a power distribution network and the access position of a distributed power supply, forming an adjacent matrix by the connection relationship among the nodes of the power distribution network, and setting a data label for the operation state of the power distribution network so as to construct a graph data set; constructing a data feature mining model for the graph data set based on the graph convolutional neural network layer; setting a loss function trained by the data feature mining model, and adjusting the hyper-parameter optimization data feature mining model; and training the data feature mining model to complete the online intelligent analysis of the novel power distribution network accident. According to the invention, through the data feature mining model based on the graph convolutional neural network, the connection relationship between the nodes is considered while the operation data of the nodes are analyzed, so that more rapid, more accurate and more flexible online accident analysis of the novel power distribution network can be realized.

Description

Technical Field

[0001] The present invention belongs to the technical field of online intelligent analysis of distribution networks, and relates to a novel online accident analysis method and system for distribution networks based on graph convolutional neural networks. Background Art

[0002] With the increase in the penetration rate of renewable energy power generation, distributed power sources, energy storage, new energy vehicles and other equipment are connected to the distribution network on a large scale, and various renewable energy power generation systems have an increasing impact on the operation and control of the distribution network. The volatility and anti-peak characteristics of renewable energy output power have a certain impact on the safe and stable operation of the power grid, and the larger the scale of renewable energy power generation, the greater the risk of safe and stable operation of the power grid.

[0003] The access of a high proportion of distributed power sources means that the distribution network contains a large number of power electronic devices, and the distribution network changes from a traditional unidirectional power flow system to a bidirectional power flow system, so the system becomes more complex.

[0004] Traditional distribution network system accident analysis methods are no longer applicable to the analysis of new distribution networks. Summary of the invention

[0005] In view of the problems existing in the above-mentioned prior art, the present invention provides a new distribution network online accident analysis method based on graph convolutional neural network, which converts the real-time operation data of the new distribution network into graph data, and inputs the real-time graph data into a data feature mining model based on graph convolutional neural network to obtain the operating status of the distribution network, thereby realizing online accident analysis of the new distribution network for high proportion of distributed power supply access.

[0006] To achieve the above object, the present invention adopts the following technical solution.

[0007] In a first aspect, the present invention provides a novel distribution network online accident analysis method based on graph convolutional neural network, which comprises the steps of:

[0008] S1, collects multi-state operation data of the new distribution network, forms a node feature data matrix with the voltage, current, power and distributed power access location of the distribution network nodes, forms an adjacency matrix with the topological relationship of the distribution network electrical nodes, sets data labels for the distribution network operation status, and constructs a graph data set with the node feature data matrix, adjacency matrix and data label values;

[0009] S2, based on the graph convolutional neural network layer, builds a data feature mining model for graph datasets;

[0010] S3, setting a loss function for training the data feature mining model, and adjusting hyperparameters to optimize the data feature mining model;

[0011] S4, training data feature mining model to complete online intelligent analysis of new distribution network accidents.

[0012] Further, in the step S1, the multiple operating states of the new distribution network refer to the operating state of the distribution network during a line short - circuit fault accident that may occur on each distribution network line and the operating state when the distribution network is normal.

[0013] Further, in the step S1, the graph data is a data structure representing entities and relationships with points and edges, as shown in Figure G:

[0014] G=(V, E) (1)

[0015] Among them, V represents the set of all points in graph G, and E represents the set of all edges in graph G. The so - called edges are the connection relationships between points.

[0016] Further, in the step S1, the node feature data matrix is as shown in Equation (2):

[0017]

[0018] Among them, A represents the node feature data matrix of the distribution network at a certain moment, and U na , U nb , U nc , I na , I nb , I nc respectively represent the amplitudes of the three - phase voltage and three - phase current of the nth node. P n , Q n respectively represent the active power and reactive power of the nth node. d n represents whether a distributed power source is connected to the nth node. d n = 1 indicates that a distributed power source is connected, and d n = 0 indicates that no distributed power source is connected.

[0019] Further, in the step S1, the adjacency matrix B is as shown in Equation (3):

[0020]

[0021] Among them, B represents the adjacency matrix of the distribution network at a certain moment. b ij = b ji = 1 indicates that there is a connection relationship between node i and node j, and b ij = b ji = 0 indicates that there is no connection relationship between node i and node j, where 0 < i < n + 1 and 0 < j < n + 1. Therefore, the adjacency matrix covers the topological structure information of the distribution network.

[0022] Furthermore, in step S1, the data label is c, c=i indicates that a short circuit fault occurs in the i-th line, c=0 indicates that the distribution network is in normal operation, and 0 <i<n+1。

[0023] Furthermore, in step S2, the convolution formula of the graph convolutional neural network layer is as follows:

[0024] g*f=U(U T g·U T f) (4)

[0025] Among them, g represents the convolution kernel function of the graph convolutional neural network layer, f represents the input node feature data matrix, and U represents the eigenvector of the Laplacian matrix of the graph. The Laplacian matrix is ​​composed of the degree matrix and adjacency matrix of the node. The degree matrix is ​​a diagonal matrix, and the elements on its diagonal are the degree of node i (that is, the number or weight sum of the edges connected to node i). The Laplacian matrix L is defined as L=DB, where D is the degree matrix and B represents the adjacency matrix of the distribution network at a certain moment.

[0026] Furthermore, in step S2, the data feature mining model includes three layers of graph convolutional neural network layers, and each graph convolutional neural network layer is connected by an activation function layer. The activation function layer increases the nonlinear characteristics of the graph convolutional neural network layer to prevent gradient disappearance and accelerate the training efficiency; the last layer uses a fully connected layer to discriminate the data features extracted by the graph convolutional neural network to realize online accident analysis of the new distribution network.

[0027] Furthermore, in step S3, a cross entropy loss function is used as a loss function for training the data feature mining model, and the function is as shown in formula (5):

[0028]

[0029] Among them, M represents the number of sample categories, y k represents the label of the k-class sample, p k Represents the predicted probability of k-type samples;

[0030] On the basis of constructing the data feature mining model and loss function, hyperparameter adjustment is performed to optimize the performance of the data feature mining model, setting the learning rate I = 0.0005, the number of hidden layer channels hidden_channels = 64, and the number of batch training samples batch_size = 64.

[0031] Furthermore, in step 4, the round at which the data feature mining model stops training is determined based on the convergence of the loss function. When the loss function is no lower than the minimum loss function for 50 consecutive rounds of training, the data feature mining model stops training and the model parameters at the minimum loss function value are saved.

[0032] In a second aspect, the present invention provides a novel distribution network online accident analysis system based on graph convolutional neural network, which comprises:

[0033] Graph data set construction unit: collects multi-state operation data of the new distribution network, forms a node feature data matrix with the voltage, current, power and distributed power access position of the distribution network nodes, forms an adjacency matrix with the topological relationship of the distribution network electrical nodes, sets data labels for the distribution network operation status, and constructs a graph data set with the node feature data matrix, adjacency matrix and data label values;

[0034] Data feature mining model building unit: Based on the graph convolutional neural network layer, a data feature mining model for graph data sets is built;

[0035] Data feature mining model optimization unit: setting the loss function of data feature mining model training, adjusting hyperparameters to optimize the data feature mining model;

[0036] Data feature mining model training unit: train data feature mining models to complete online intelligent analysis of new distribution network accidents.

[0037] The present invention has the beneficial effects: Compared with the prior art, the present invention provides a new distribution network online accident analysis method based on graph convolutional neural network, which mines the node feature data of the new distribution network through the deep learning algorithm graph convolutional neural network while considering the distribution network topology information, thereby realizing a more accurate and better generalized new distribution network online accident analysis method. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, for ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0039] Figure 1 It is a flow chart of a novel distribution network online accident analysis method based on graph convolutional neural network of the present invention;

[0040] Figure 2 It is a structural diagram of a data feature mining model based on a graph convolutional neural network of the present invention;

[0041] Figure 3It is a topological diagram of the IEEE33 node distribution network of the distributed power source in the application example of the present invention;

[0042] Figure 4 It is a curve diagram of the loss function of the data feature mining model training based on the graph convolutional neural network in the application example of the present invention;

[0043] Figure 5 This is a structural diagram of a new distribution network online accident analysis system based on graph convolutional neural network in the present invention. DETAILED DESCRIPTION

[0044] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0045] Example 1

[0046] This embodiment provides a new distribution network online accident analysis method based on graph convolutional neural network. Figure 1 As shown, the steps are as follows:

[0047] S1, collects multi-state operation data of the new distribution network, forms a node feature data matrix with the voltage, current, power and distributed power access location of the distribution network nodes, forms an adjacency matrix with the topological relationship of the distribution network electrical nodes, sets data labels for the distribution network operation status, and constructs a graph data set with the node feature data matrix, adjacency matrix and data label values;

[0048] S2, based on the graph convolutional neural network layer, builds a data feature mining model for graph datasets;

[0049] S3, setting a loss function for training the data feature mining model, and adjusting hyperparameters to optimize the data feature mining model;

[0050] S4, training data feature mining model to complete online intelligent analysis of new distribution network accidents.

[0051] In step S1, the various operating states of the new distribution network refer to the operating states of the distribution network when a line short circuit fault accident may occur on each distribution network line and the operating states of the distribution network when it is normal.

[0052] In step S1, graph data is a data structure that represents entities and relationships using nodes and edges, as shown in Figure G:

[0053] G=(V,E) (1)

[0054] Among them, V represents the set of all points in the graph G, E represents the set of all edges in the graph G, and the edges are the connection relationships between the points.

[0055] In step S1, the node feature data matrix is ​​as shown in formula (2):

[0056]

[0057] Among them, A represents the node characteristic data matrix of the distribution network at a certain moment, U na , U nb , U nc ,I na ,I nb ,I nc Respectively represent the three-phase voltage and three-phase current amplitudes of the nth node, P n , Q n Represent the active power and reactive power of the nth node respectively, d n Indicates whether the nth node is connected to a distributed power source, d n =1 means that distributed power is connected, d n =0 means no distributed power generation is connected.

[0058] In step S1, the adjacency matrix B is as shown in formula (3):

[0059]

[0060] Among them, B represents the adjacency matrix of the distribution network at a certain moment, b ij =b ji =1 means that node i and node j are connected, b ij =b ji =0 means that node i has no connection with node j, 0 <i<n+1,0<j<n+1。

[0061] In step S1, the data label is c, where c=i indicates that a short circuit fault occurs in the i-th line, c=0 indicates that the distribution network is in normal operation, and 0 <i<n+1。

[0062] In step S2, the convolution formula of the graph convolutional neural network layer is as follows:

[0063] g*f=U(U T g·U T f) (4)

[0064] Among them, g represents the convolution kernel function of the graph convolutional neural network layer, f represents the input node feature data matrix, and U represents the eigenvector of the Laplacian matrix of the graph. The Laplacian matrix is ​​composed of the degree matrix and adjacency matrix of the node. The degree matrix is ​​a diagonal matrix, and the elements on its diagonal are the degree of node i (that is, the number or weight sum of the edges connected to node i). The Laplacian matrix L is defined as L=DB, where D is the degree matrix and B represents the adjacency matrix of the distribution network at a certain moment.

[0065] In step S2, Figure 2 As shown, the data feature mining model includes three layers of graph convolutional neural network layers, and each graph convolutional neural network layer is connected by an activation function layer. The activation function layer increases the nonlinear characteristics of the graph convolutional neural network layer to prevent the gradient from disappearing and accelerate the training efficiency; the last layer uses a fully connected layer to discriminate the data features extracted by the graph convolutional neural network layer to realize online accident analysis of the new distribution network.

[0066] In step S3, a cross entropy loss function is used as a loss function for training the data feature mining model. The function is as shown in formula (5):

[0067]

[0068] Among them, M represents the number of sample categories, y k represents the label of the k-class sample, p k Represents the predicted probability of k-type samples;

[0069] On the basis of building a data feature mining model and loss function, we adjust the hyperparameters to optimize the performance of the data feature mining model, setting the learning rate I = 0.0005, the number of hidden layer channels hidden_channels = 64, and the number of batch training samples batch_size = 64. Among them, the learning rate determines the size of the network weight adjustment during the training process of the data feature mining model. The larger the learning rate, the faster the training speed, but it is easy to skip the global minimum of the loss function; the number of hidden layer channels directly affects the computational complexity and performance of the data feature mining model. The more channels, the higher the accuracy of the data feature mining model, but the computational complexity will also increase greatly; the number of batch training samples affects the training speed and generalization ability of the data feature mining model. The larger the number, the faster the training speed and the faster the convergence, but the more memory is occupied, the worse the generalization ability of the data feature mining model.

[0070] In step 4, the round at which the data feature mining model based on the graph convolutional neural network stops training is determined according to the convergence of the loss function. When the loss function is no lower than the minimum loss function after 50 consecutive rounds of training, the data feature mining model stops training and the model parameters at the minimum loss function value are saved.

[0071] Application Examples

[0072] The method described in Example 1 was used as follows.

[0073] The network in this application example adopts the IEEE33 node distribution network model, and its network structure diagram is as follows: Figure 3 As shown in the figure, the system is connected to distributed power sources at distribution network nodes 18, 22, 25, and 33 respectively. The IEEE 33-node distribution network system is built in matlab / simulink, and the script file is used to batch generate multi-state operation data of the distribution network. The distribution network operation state sets 33 operation states for each line, including short-circuit fault and normal operation. Sample data within 0.5 seconds before and after the fault occurs, and 300 sampling points are taken for each branch, totaling 9,600 data. The data set is shuffled and divided into training set, validation set, and test set according to 8:1:1.

[0074] like Figure 4 This is the curve of the loss function and accuracy of the data mining model training based on graph convolutional neural network as the training rounds epoch change. It can be seen from the figure that with the increase of training rounds, the loss function continues to decrease and the accuracy continues to increase. When the number of iterations exceeds 500 rounds, the loss function converges to the accuracy, and the accuracy reaches 98.82%.

[0075] After the model training is completed, the divided test set data is used to verify the effect of the model. The evaluation indicators are analyzed from the aspects of precision, recall rate, and F1 score. The number of samples in each category in the test set is not exactly the same. Therefore, macro average and weighted average are used here to describe the three evaluation indicators as shown in Table 1:

[0076] Table 1

[0077] Evaluation indicators Accuracy / % Recall / % F1 score / % Macro Average 98.71 98.77 98.70 Weighted Average 98.80 98.71 98.72

[0078] The data in Table 1 show that the recognition accuracy of the data mining model is higher than 98.7% in all descriptive aspects, and has a high recognition accuracy.

[0079] In summary, this application example confirms the feasibility of a new distribution network online accident analysis method based on graph convolutional neural network of the present invention. By constructing a graph data including a node data feature matrix of distribution network operation data and an adjacency matrix including distribution network topology information, a data mining model based on graph convolutional neural network is used to perform data feature mining on distribution network operation data to realize a new distribution network online accident analysis.

[0080] Example 2

[0081] This embodiment provides a new distribution network online accident analysis system based on graph convolutional neural network. Figure 5 As shown, it includes:

[0082] Graph data set construction unit: collects multi-state operation data of the new distribution network, forms a node feature data matrix with the voltage, current, power and distributed power access position of the distribution network nodes, forms an adjacency matrix with the topological relationship of the distribution network electrical nodes, sets data labels for the distribution network operation status, and constructs a graph data set with the node feature data matrix, adjacency matrix and data label values;

[0083] Data feature mining model building unit: Based on the graph convolutional neural network layer, a data feature mining model for graph data sets is built;

[0084] Data feature mining model optimization unit: setting the loss function of data feature mining model training, adjusting hyperparameters to optimize the data feature mining model;

[0085] Data feature mining model training unit: train data feature mining models to complete online intelligent analysis of new distribution network accidents.

[0086] It should be noted that each unit in the above-mentioned new distribution network online accident analysis system based on graph convolutional neural network can be fully or partially implemented by software, hardware and their combination. The above-mentioned units can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory in the computer device in the form of software, so that the processor can call and execute the operations corresponding to the above-mentioned units. For the specific definition of a new distribution network online accident analysis system based on graph convolutional neural network, please refer to the above definition of a new distribution network online accident analysis method based on graph convolutional neural network (ie, Example 1). The two have the same functions and effects, which will not be repeated here.

[0087] The above disclosure is only the preferred embodiment of the present invention, showing and describing the basic principle, main features and advantages of the present invention. However, the present invention is not limited thereto, and any transformation and changes made by any technician in the field without creative work and without departing from the principle of the present invention should fall within the protection scope of the present invention. The protection scope of the present invention is defined by the attached claims and their equivalents.

Claims

1. A new distribution network online accident analysis method based on graph convolutional neural network, characterized in that: Includes steps: S1, collects multi-state operation data of the new distribution network, forms a node feature data matrix with the voltage, current, power and distributed power access location of the distribution network nodes, forms an adjacency matrix with the topological relationship of the distribution network electrical nodes, sets data labels for the distribution network operation status, and constructs a graph data set with the node feature data matrix, adjacency matrix and data label values; S2, based on the graph convolutional neural network layer, builds a data feature mining model for graph datasets; S3, setting a loss function for training the data feature mining model, and adjusting hyperparameters to optimize the data feature mining model; S4, training the data feature mining model to complete the online intelligent analysis of new distribution network accidents.

2. The novel distribution network online accident analysis method according to claim 1, characterized in that: In step S1, graph data is a data structure that represents entities and relationships using nodes and edges, as shown in Figure G: G=(V,E) (1) Among them, V represents the set of all points in the graph G, E represents the set of all edges in the graph G, and the edges are the connection relationships between the points.

3. The novel distribution network online accident analysis method according to claim 1, characterized in that: In step S1, the node feature data matrix is ​​as shown in formula (2): Among them, A represents the node characteristic data matrix of the distribution network at a certain moment, U na , U nb , U nc ,I na ,I nb ,I nc Respectively represent the three-phase voltage and three-phase current amplitudes of the nth node, P n , Q n Represent the active power and reactive power of the nth node respectively, d n Indicates whether the nth node is connected to a distributed power source, d n =1 means that distributed power is connected, d n =0 means no distributed power generation is connected.

4. The novel distribution network online accident analysis method according to claim 1, characterized in that: In step S1, the adjacency matrix B is as shown in formula (3): Among them, B represents the adjacency matrix of the distribution network at a certain moment, b ij =b ji =1 indicates that node i and node j are connected, b ij =b ji =0 means that node i has no connection with node j, 0 <i<n+1,0<j<n+1。 5. The novel distribution network online accident analysis method according to claim 1, characterized in that: In step S1, the data label is c, where c=i indicates that a short circuit fault occurs in the i-th line, c=0 indicates that the distribution network is in normal operation, and 0 <i<n+1。 6. The novel distribution network online accident analysis method according to claim 1, characterized in that: In step S2, the convolution formula of the graph convolutional neural network layer is as follows: g*f=U(U T g·U T f) (4) Among them, g represents the convolution kernel function of the graph convolutional neural network layer, f represents the input node feature data matrix, and U represents the eigenvector of the Laplace matrix of the graph; the Laplace matrix is ​​composed of the degree matrix and adjacency matrix of the node. The degree matrix is ​​a diagonal matrix, and the elements on its diagonal are the degree of node i. The Laplace matrix L is defined as L=DB, where D is the degree matrix and B represents the adjacency matrix of the distribution network at a moment.

7. The novel distribution network online accident analysis method according to claim 1, characterized in that: In step S2, the data feature mining model includes three layers of graph convolutional neural network layers, and each graph convolutional neural network layer is connected by an activation function layer. The activation function layer increases the nonlinear characteristics of the graph convolutional neural network layer to prevent the gradient from disappearing and accelerate the training efficiency; the last layer uses a fully connected layer to discriminate the data features extracted by the graph convolutional neural network layer to realize online accident analysis of the new distribution network.

8. The novel distribution network online accident analysis method according to claim 1, characterized in that: In step S3, a cross entropy loss function is used as a loss function for training the data feature mining model. The function is as shown in formula (5): Among them, M represents the number of sample categories, y k represents the label of the k-class sample, p k Represents the predicted probability of k-type samples; On the basis of constructing the data feature mining model and loss function, hyperparameter adjustment is performed to optimize the performance of the data feature mining model, setting the learning rate I = 0.0005, the number of hidden layer channels hidden_channels = 64, and the number of batch training samples batch_size = 64.

9. The novel distribution network online accident analysis method according to claim 1, characterized in that: In step 4, the round at which the data feature mining model stops training is determined according to the convergence of the loss function. When the loss function is no lower than the minimum loss function after 50 consecutive rounds of training, the data feature mining model stops training and the model parameters at the minimum loss function value are saved.

10. A new distribution network online accident analysis system based on graph convolutional neural network, characterized in that: include: Graph data set construction unit: collects multi-state operation data of the new distribution network, forms a node feature data matrix with the voltage, current, power and distributed power access position of the distribution network nodes, forms an adjacency matrix with the topological relationship of the distribution network electrical nodes, sets data labels for the distribution network operation status, and constructs a graph data set with the node feature data matrix, adjacency matrix and data label values; Data feature mining model building unit: Based on the graph convolutional neural network layer, a data feature mining model for graph data sets is built; Data feature mining model optimization unit: setting the loss function of data feature mining model training, adjusting hyperparameters to optimize the data feature mining model; Data feature mining model training unit: train data feature mining models to complete online intelligent analysis of new distribution network accidents.

Images