High-proportion new energy power distribution network open-closed loop state identification method based on GCNN

Abstract

The invention discloses a high-proportion new energy power distribution network open-closed loop state identification method based on a GCNN, and belongs to the technical field of calculation, calculation or counting. The system is composed of operation state sensors at all new energy nodes, power detection devices between links, an upper computer used for data processing and state recognition and corresponding communication links. According to the method, a graph network is constructed for a target power distribution network topology, and then a weighted attribute graph data set is constructed by combining node state information and inter-link power information of the target power distribution network. And training the graph convolution pooling neural network based on the weighted attribute graph data set, and finally identifying the open-loop and closed-loop states of the power distribution network by adopting the trained graph neural network. According to the method, the problems that in the high-proportion new energy power distribution network, switch tripping is caused by a non-automatic switch, and open and closed loop states are difficult to detect are solved, risks can be found in time, and the operation reliability and safety of the high-proportion new energy power distribution network are effectively improved.

Description

Technical Field

[0001] The invention relates to power system protection technology, and discloses a method for identifying open-loop and closed-loop states of a high-proportion new energy distribution network based on a graph convolutional neural network (GCNN), which belongs to the technical field of calculation, extrapolation or counting. Background Art

[0002] As one of the means to optimize the operation of distribution networks, distribution network reconstruction technology changes the topological structure of distribution networks to achieve the goals of reducing network losses, improving voltage distribution, and improving system reliability and economy. The opening and closing of distribution networks is triggered by the opening and closing actions of automated or non-automated switches. Under non-automated switches or on-site operations, it is difficult to monitor the switch status and closing and opening operations in time due to communication interruptions or topological disconnection. At the same time, the current renewable energy power generation is transforming from supplementary energy to alternative energy, and the system presents power electronics characteristics. Especially in areas with a high proportion of renewable energy, the transmission lines are long, the conventional power supply support capacity is weak, and the grid strength is low. If the open and closed loop status of a high proportion of renewable energy distribution networks cannot be accurately identified, the large voltage changes and current fluctuations caused by the corresponding operations will bring great risks to the safe and stable operation of the power grid.

[0003] The natural topological structure of the distribution network allows it to be modeled as a complex graph structure, in which nodes represent new energy equipment such as photovoltaics and wind power, and edges represent the electrical connection relationship between the equipment. For such a data form, traditional methods usually need to manually extract features when processing complex topological dependencies, which is inefficient and lacks robustness. When judging the open and closed-loop operations of the distribution network involving complex topological relationships, traditional distribution network data processing methods are difficult to meet the requirements of quickly and accurately identifying the open and closed loop states, thereby affecting the stability and safety of the power grid operation. Graph Neural Network (GNN) is a neural network model specially designed for graph structure data. It can effectively capture the dependency relationship between nodes and edges in the graph by iteratively aggregating nodes and their neighborhood information. This makes GNN show unique advantages in the processing and analysis of power grid data. The present invention aims to propose a method for identifying the open and closed loop states of a high-proportion new energy distribution network based on GNN to overcome the defects of existing distribution network data processing methods. Summary of the invention

[0004] The purpose of the present invention is to address the deficiencies of the above-mentioned background technology and to provide a method for identifying the open- and closed-loop states of a high-proportion new energy distribution network based on a graph convolutional neural network, so as to solve the technical problem that traditional distribution network data processing methods are difficult to quickly and accurately identify the open- and closed-loop states of a high-proportion new energy distribution system, and to achieve the purpose of efficiently processing graph structure data in the distribution network and accurately identifying the risks that may be caused by open- and closed-loop operations.

[0005] The present invention adopts the following technical solutions to achieve the above-mentioned invention object:

[0006] The open-loop and closed-loop state recognition method of a high-proportion renewable energy distribution network based on GCNN includes the following steps:

[0007] Step 1: Map the node devices of the target high-proportion new energy distribution network and the electrical connection relationship between the node devices into vertex sets and edge sets respectively, construct graph data including vertex sets, edge sets, weight matrices and vertex attribute matrices, preprocess the node data of the target high-proportion new energy distribution network, assign the preprocessed node data of the target high-proportion new energy distribution network to the graph data, and obtain a labeled weighted attribute graph data set;

[0008] Step 2: Use a labeled weighted attribute graph dataset to train a graph convolutional neural network that includes at least one convolutional layer and a prediction layer. Each convolutional layer performs convolution and pooling operations on the input graph data to obtain the graph data and the coarsened graph after aggregating the attributes of the k-order neighbor vertices. The prediction layer normalizes the coarsened graph output by the last convolutional layer to obtain the predicted label that characterizes the open-loop and closed-loop state of the target high-proportion new energy distribution network. Back-propagate the error between the predicted label and the true label, and update the graph convolutional neural network parameters.

[0009] Step 3: Collect node data of the target high-proportion new energy distribution network, obtain a weighted attribute graph data set, input the weighted attribute graph data set into the trained graph convolutional neural network, and obtain the recognition results of the open-loop and closed-loop states of the target high-proportion new energy distribution network.

[0010] As a further optimization scheme for the open-and-closed-loop state identification method of a high-proportion new energy distribution network based on GCNN, in step 1, the vertex attribute matrix includes the attribute data of each node device, and the attribute data of each node device includes but is not limited to: specific operating data and text information. The specific operating data includes but is not limited to: temperature and humidity, voltage, and current. The text information includes but is not limited to: normal temperature and high temperature.

[0011] As a further optimization scheme of the open-and-closed-loop state identification method of a high-proportion new energy distribution network based on GCNN, in step 1, the edge set includes but is not limited to the following information: whether the node devices of the target high-proportion new energy distribution network are topologically connected and the edge weight of the topological connection between the node devices. The edge weight of the topological connection between the node devices is represented by the transmission power information between the node devices.

[0012] As a further optimization scheme of the open-and-closed-loop state identification method of a high-proportion renewable energy distribution network based on GCNN, in step 2, the k-order neighbor vertices are the vertices corresponding to the neighbor nodes whose number of nodes separated by the shortest topological connection between the nodes corresponding to each vertex is k.

[0013] As a further optimization scheme for the open-and-closed-loop state recognition method of a high-proportion renewable energy distribution network based on GCNN, in step 2, the acquisition of the aggregated k-order neighbor vertex attributes is achieved by averaging the aggregated k-order neighbor vertex attribute vectors. Among them, N k (v i ) is the i-th vertex v i The vertices in the k-order neighborhood of |N k (v i )| is the number of k-order neighbor nodes of the i-th vertex, The vector after the attributes of the k-order neighbor vertices of the i-th vertex are aggregated. is the jth neighbor vertex v j The attribute vector, w ij For the i-th vertex v i and the jth neighbor vertex v j The weight between ij By i and the jth neighbor vertex v j The edge weights of multiple connected topological lines are summed up and obtained.

[0014] As a further optimization scheme of the open-and-closed-loop state recognition method of a high-proportion new energy distribution network based on GCNN, in step 2, the vertex attribute matrix of the graph data input to the current convolutional layer is composed of the vector after the attributes of the k-order neighbor vertices of the previous convolutional layer are aggregated and the vertex attribute matrix of the graph data output by the current convolutional layer.

[0015] As a further optimization scheme of the open-and-closed-loop state recognition method for a high-proportion renewable energy distribution network based on GCNN, in step 2, the specific method by which each convolutional layer performs pooling operation on the input graph data is as follows: according to the information gain of the vertex set of the graph data input to the current convolutional layer, the graph data after the current convolutional layer aggregates the attributes of the k-order neighbor vertices is downsampled.

[0016] As a further optimization scheme for the open and closed-loop state recognition method of a high-proportion renewable energy distribution network based on GCNN, the specific method for downsampling the graph data after the current convolutional layer aggregates the attributes of k-order neighbor vertices is: traverse the vertex set of the graph data after the current convolutional layer aggregates the attributes of k-order neighbor vertices, and prune the vertices whose information gain is less than the threshold.

[0017] As a further optimization scheme of the open-loop and closed-loop state recognition method of high-proportion renewable energy distribution network based on GCNN, according to the expression

[0018] Calculate the information gain of the vertex, where IG(v j ) is the jth vertex v j The information gain, P(y s ) is the label y s The probability, P(y s |v j ) is based on the j-th vertex v j Attribute data classifies graph data into labels y s probability.

[0019] The open-closed loop state identification system of a high-proportion new energy distribution network based on GCNN includes: an operation state sensor at each new energy node, a power detection device between each link, and a host computer for data processing and state identification; the host computer receives data collected by the operation state sensor and the power detection device, and executes the above method.

[0020] The present invention adopts the above technical solution and has the following beneficial effects:

[0021] (1) The GCNN-based open-loop and closed-loop state recognition method for a high-proportion renewable energy distribution network proposed in the present invention first identifies and distinguishes high- and low-level attributes of nodes in the distribution network by aggregating attribute vectors, highlighting attribute data that is more representative of the topology judgment of the distribution network; then, the graph is coarsened based on information gain to highlight the correlation between different nodes in the power grid. This method can effectively identify the topological structure of a high-proportion renewable energy distribution network that uses power grid reconstruction technology, and solve the problem of switch tripping and difficulty in detecting open-loop and closed-loop states caused by non-automated switches.

[0022] (2) The GCNN-based open-loop and closed-loop state recognition method for a high-proportion new energy distribution network proposed in the present invention assigns the power information between node devices to the edge attributes in the graph data when constructing graph data. In the process of aggregating attribute vectors, the edge weights of the target vertex and neighbor node connection topology are used to assign the weights of adjacent nodes, which can effectively filter out low-level attribute data. Accompanied by alternating convolution operations and pooling operations, it can effectively identify high-level attribute data, providing a basis for the graph neural network to accurately capture key attribute data.

[0023] (3) The GCNN-based open-loop and closed-loop state identification method for a high-proportion new energy distribution network proposed in the present invention discovers vertices that can accurately classify graph data by screening vertices whose information gain is less than a threshold, optimizes graph data in combination with pruning operations, efficiently processes graph structure data in the distribution network, accurately identifies the risks that may be caused by open-loop and closed-loop operations, and ensures the stability and safety of grid operation. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 This is a schematic diagram of a method for detecting open-loop and closed-loop states of a high-proportion new energy distribution network based on a graph convolutional neural network proposed in the present invention.

[0025] Figure 2 This is a schematic diagram of the open and closed loop of a high-proportion renewable energy distribution network.

[0026] Figure 3 It is a weighted attribute graph constructed for the topology of distribution network with high proportion of renewable energy.

[0027] Figure 4 This is a schematic diagram of the forward propagation principle of the graph convolutional neural network proposed in the present invention. DETAILED DESCRIPTION

[0028] The technical solution of the invention is described in detail below with reference to the accompanying drawings.

[0029] like Figure 1 As shown, the open-loop and closed-loop state detection method of a high-proportion new energy distribution network based on a graph convolutional neural network proposed in the present invention is mainly divided into two parts, a data processing part and a distribution network open-loop and closed-loop state recognition part. Among them, in the data processing part, the distribution network node data and the distribution network graph data are preprocessed to obtain a weighted attribute graph data set; the distribution network open-loop and closed-loop state recognition part is implemented by a graph convolutional neural network, which includes a convolutional layer, a pooling layer and a prediction layer. The number of convolutional layers and pooling layers depends on the specified neural network depth. After the weighted attribute graph data set is processed by convolution and pooling, a coarsened graph is obtained. The prediction layer normalizes the coarsened graph to obtain the recognition results of the open-loop and closed-loop states of the target distribution network.

[0030] In the present invention, the graph convolutional neural network can achieve the above functions after the training stage and the inference stage. In the training stage, the graph convolutional neural network model proposed in the present invention finds the optimal parameters by training a labeled weighted attribute graph data set. The parameters of the graph convolutional neural network model proposed in the present invention include the filter vector F, the number of filters f and the weight matrix W. The open-loop and closed-loop state recognition results of the target distribution network are obtained by the forward propagation of the convolutional pooling algorithm, and the detailed process of the forward propagation of the convolutional pooling algorithm will be given later. The error of predicting the open-loop and closed-loop states of the target distribution network is then back-propagated to adjust the parameters until convergence.

[0031] In the inference stage, the classification task of the graph convolutional neural network model proposed in this invention is to predict the labels of the coarsened graph, which are unknown.

[0032] Assume that the training data set is S t ={(G1,y1),(G2,y2),(G3,y3),...,(G t ,y t )}, where t is the number of training samples, G is a weighted attributed graph, and Y is a label set containing S labels Y = {y1, y2, y3, ..., y S}, while the test data set is a graph data set without labels. The node devices in the target distribution network are mapped as vertices, and the electrical connections between the node devices in the target distribution network are mapped as edges. Based on the four graph structure parameters V, E, W, and X, the graph network mapped by the distribution network can be characterized. Let G = (V, E, W, X) be a graph network with unknown labels, where V is the vertex set, E is the edge set, W is the weight matrix, and X is the vertex attribute matrix.

[0033] The edge set indicates whether the target distribution network node devices are topologically connected and the edge weight of the topological connection between the node devices, where the edge weight of the topological connection between the node devices is represented by the transmission power information between the node devices. Considering that the topological connection status between the node devices in the target distribution network is symmetrical, the edge set E can be represented by a symmetric matrix with all diagonals being 0. Taking the target distribution network containing 4 node devices as an example, E is shown in formula (1.1). The 1234 on the top and left side of the matrix represent the node device numbers, and the corresponding rows and columns of the matrix represent the topological connection relationship between the node devices corresponding to the numbers. The value of each element in the matrix is ​​the transmission power between the node devices corresponding to the numbers. If the element value is 0, it means there is no connection.

[0034]

[0035] Under different requirements, the specific values ​​of each element in the weight matrix W depend on the attribute weights of the node devices in the target power distribution network. The vertex attribute matrix includes the attribute data of each node device, and the attribute data of each node device includes: specific operating data such as temperature and humidity, voltage, current, or text information such as normal temperature and high temperature. The goal of the graph convolutional neural network model proposed in this invention is to derive the mapping function f:G→Y, which predicts the category label y of a given graph G. s .

[0036] The present invention accomplishes this task by designing a convolution layer, which is symbolically represented by Convolution(). The convolution layer updates the attribute representation of each vertex by aggregating the attributes of the k-order neighbor vertices of each vertex. The k-order neighbor vertices represent the vertices corresponding to the neighbor nodes whose number of nodes separated by the shortest topological connection between the nodes corresponding to each vertex is k. For example, node devices A, B, and C form a topological connection ABC. If there is no direct connection relationship between A and C, A and C are 2nd-order neighbor nodes, and A and B, B and C are 1st-order neighbor nodes.

[0037] The output of the convolutional layer is a graph G with a new vertex attribute representation zv The output of the convolutional layer is then passed to the pooling layer, which is symbolized as Pooling(). In the pooling layer, the goal is to find a coarsened graph G C =(V C ,E C ),in, is the vertex set of the coarsened graph, is the edge set of the coarsened graph, and |V C |<<|V zv | and |E C |<<|E|.

[0038] The prediction generation layer is a typical fully connected layer that only contains a feedforward neural network. The prediction generation layer is symbolically represented as Predict(). The detailed working principle of the convolutional layer will be given in the next section, followed by the pooling layer and the prediction generation layer. Algorithm 1 shows the workflow of the graph convolutional neural network model proposed in this invention.

[0039] Algorithm 1

[0040] Input parameters: graph structure data G; vertex attribute matrix Represents the attribute of each vertex, n is the number of vertices, d is the attribute dimension. The number of layers of the neural network is l.

[0041] Output parameter: label y of the graph s , in the present invention, the open-loop and closed-loop status labels of the distribution network

[0042] Algorithm parameters: filter F, used for convolution operation, also called convolution kernel; number of filters in the convolution layer f; weight matrix W of layer l l , used for linear transformation.

[0043] 1. Cycle operation:

[0044] For each layer i from 1 to l:

[0045] 1.1. Convolution operation:

[0046] Apply the convolution layer to convolve the input graph G to generate a new attribute representation

[0047] 1.2. Pooling operation:

[0048] After convolution, Perform pooling to generate a coarsened sub-graph

[0049] 2. End the loop.

[0050] 3. Forecast Generation:

[0051] Use the prediction generation layer to coarsen the subgraph G of the last layer l C Make a prediction and get the label y of graph G s .

[0052] 4. Return the open and closed loop judgment results of the distribution network y s

[0053] The following will provide a detailed introduction in conjunction with the specific steps in the content of the invention.

[0054] Step 1: construct a labeled weighted attribute graph dataset for the target high-proportion new energy distribution network;

[0055] Step 1.1, construct corresponding graph data based on the target distribution network topology; Figure 2 The diagram shown is a schematic diagram of a multi-node distribution network architecture, which is for illustrative purposes only, and the present invention is not limited to this distribution network. Figure 2 The multi-node distribution network listed includes four main nodes and multiple secondary nodes. The main nodes are AC grid access nodes, wind power nodes, photovoltaic nodes and electrochemical energy storage nodes. The secondary nodes can perform operations such as distribution terminal connection, which will not be described in detail in the present invention. Figure 2 The solid lines in the figure are branches, and the dotted lines are tie switches. The topology of the distribution network can be changed through distribution network reconstruction technology to reduce network losses, improve voltage distribution, and enhance system reliability and economy. Figure 2The distribution network topology shown in the figure can be simplified to the following without performing any distribution network reconstruction operations such as closing the tie switch. Figure 3 The network diagram shown.

[0056] Step 1.2, collecting data of each node and edge in the target high-proportion new energy distribution network graph data, specifically including attribute data and edge data of each node; Figure 3 The network shown contains four nodes and five topological connection edges, which can form a graph data containing four vertices and five edges. The distribution network operator collects node attribute data, collects k groups of attribute data for each node, and collects edge data between nodes, where the edge data is the transmission power between distribution network nodes.

[0057] In step 1.3, the collected node attribute data and edge data are preprocessed to convert the data into real values ​​rather than categorical or text variables.

[0058] Step 1.4, assign the preprocessed node attribute data and edge data to the graph network data constructed in step 1.1 to obtain a labeled weighted attribute graph dataset.

[0059] Step 2: Train the graph convolutional neural network based on the labeled weighted attribute graph dataset to construct Figure 4 The graph convolutional neural network forward transmission framework shown in the figure is used to describe the graph convolutional neural network forward propagation proposed in the present invention.

[0060] The graph convolutional neural network receives a graph as input and generates an output label, which in this invention is the open-loop and closed-loop state label of the distribution network operation. The output label result is then subtracted from the original label result to obtain the error, and then the recognition error is back-propagated to adjust the parameters of the graph convolutional pooling neural network to minimize the cost or error. Below, a detailed introduction will be given in conjunction with the specific steps.

[0061] Step 2.1: any labeled graph data is input into the filter, attribute data of the input graph data is aggregated, and the graph data is updated;

[0062] In the present invention, a labeled weighted attribute graph dataset consisting of distribution network data is used as input. These graphs have a fixed topological structure and an arbitrary size. The labeled weighted attribute graph dataset is then passed to the convolution layer. The convolution layer receives the graph data, analyzes the graph, and applies a randomly initialized filter function to generate a new vertex attribute representation. By aggregating the attributes of adjacent vertices in the original input weighted attribute graph, the graph convolutional neural network model can identify and distinguish low-level attributes. The specific method of aggregating adjacent vertex attributes refers to the average converged attribute vector expression below. The graph convolutional neural network involved in the present invention includes multiple convolutional layers. As the network deepens, the convolutional layer will try to learn the high-level attributes of the graph.

[0063] Graph convolution is the core operation of the neural network model proposed in the present invention and is performed in the convolution layer. The main purpose of the convolution operation is to analyze the potential attributes of complex graph data and find useful attribute representations in order to efficiently make accurate predictions. In-depth analysis of graph data shows that close vertices have similar attributes, so they usually have the same category labels. The graph convolutional neural network proposed in the present invention learns the optimal attribute representation of each vertex in the graph data by aggregating the attributes of neighbor vertices in the k-order neighborhood of the vertex.

[0064] The following is a mathematical expression of the aggregation function: Assume that the target vertex is v3. In order to aggregate the attribute data vector of vertex v3, the present invention considers the vertex attribute data in the first-order neighborhood of v3. Let the vertices in the first-order neighborhood of v3 be represented by N1(v3), H N1 (v3) represents the average converged attribute data vector of target vertex v3, and its expression is:

[0065]

[0066] in, represents the jth neighbor vertex v j The attribute vector of ij is the target vertex v i and its adjacent vertex v j The weight between ij Specifically represents the target vertex v i and its adjacent vertex v j The edge weights of multiple connected topological lines are summed. For example, if there is an additional connection between node A and node C, the weight w of the second-order connection is considered. AC Just for E AC +(E AB +E BC ), the former E AC indicates that node A and node C are directly connected, and the latter indicates that node A and node C are connected across nodes; N1(v3) represents the set of first-order neighboring vertices of the target vertex v3; |N1(v3)| is the number of vertices in the first-order neighborhood of the target vertex v3; is the result of convergence, that is, the weighted average of the attribute vectors of vertex v3 and its first-order neighboring vertices.

[0067] The present invention uses the sum of all edge weights between two vertices as the weight between vertices, which can aggregate the connection relationship between adjacent vertices. The sum of multiple attribute vectors multiplied by weights can highlight the high-level attributes in the vertex attributes and weaken the interference of low-level attributes on topological judgment. Compared with the existing patents that require complex calculations such as mean square error, maximum value, and convolution for attribute aggregation, the present invention only emphasizes the strength of the connection relationship between nodes and the level of attributes, avoiding the situation where the accuracy of open and closed loop judgment is reduced due to information coupling introduced by additional complex operations.

[0068] The following Algorithm 2 shows the working principle of the convolutional layer of the present invention. The input of the convolutional layer of the neural network in the present invention is: a weighted and attributed graph G, the input attribute vectors of all vertices The number of layers is l, and the search depth is k; the final output is a graph with new vertex representations

[0069] The outer loop of the algorithm shows the current convolutional layer. For a given k, find the neighbor vertices u∈N of each vertex v k (v), for all v∈V. In each step of the convolutional layer, H (i) Represents the attribute vector of the vertex in the graph data output by the i-th convolutional layer of the current layer. In each layer of the neural network, for each vertex in the graph G The algorithm represents the attributes of vertices in its immediate neighborhood Aggregate() function into a single vector H (i) Note that the aggregation in this step relies on the attribute representation generated in the previous iteration i-1 of the outer loop, as well as the baseline case i = 0. When i = 0, the input vertex features are represented by the vector H (0) In the initial stage, the attribute vector x of each vertex is v Convert to H through the linear transformation function Trans() (0) .

[0070] After aggregating neighbor attribute vectors, the present invention represents the current attribute of the vertex as H (i-1) and the attribute vector H after aggregating the neighborhood (i) The concatenation is performed by the function Concatenate(). The concatenated attribute representation is then processed by a nonlinear activation function σ. In the convolution layer, the present invention propagates the attribute representation output after the activation processing to the next step H of the algorithm. (i) The present invention denotes the final attribute representation of the output of the first layer of neural network as Graph G with new representation z is the output of the convolutional layer. In order to learn useful parameters, such as the weight matrix W l(corresponding to all 1,...,l layer networks), filter vector F, and number of filters f, the present invention applies a graph-based loss function to the output representation. The present invention adjusts the designed graph convolutional pooling neural network by stochastic gradient descent to find the optimal parameters. The graph-based loss function encourages neighboring vertices to have similar representations.

[0071] Algorithm 2

[0072] Input parameters: graph G; number of layers l; search depth k.

[0073] Output parameter: graph with new vector representation

[0074] Algorithm parameters: filter F; number of filters f; weight matrix W of layer l l .

[0075] 1. Assignment

[0076] 2. For i=1 to l, perform the following operations:

[0077] 2.1. For each vertex v∈V execute:

[0078] Assignment

[0079] Assign H (i) ←σ(W (i) .Concatenate(H (i-1) ,H (i) ))

[0080] 3. End the loop

[0081] 4.

[0082] 5. Return to G zv

[0083] The above algorithm finally outputs G zv This is the final output of the convolutional layer. Next, the model applies an activation function to the output of the convolutional layer.

[0084] Step 2.2, the graph data with new attribute representation obtained by the convolution layer is input into a point-by-point nonlinear activation function, and then undergoes a pooling operation to find statistically relevant nodes in the graph network; the present invention uses a ReLU activation function, and the output graph with new attribute representation is passed to the pooling layer. The pooling layer receives the graph with new attribute representation as input.

[0085] Step 2.3, coarsen the graph network based on information gain; the pooled convolutional graph neural network proposed in the present invention finds the threshold of information gain for each graph and calculates the information gain of each vertex. If the information gain of a vertex is lower than the threshold, the proposed network model will eliminate the vertex, thereby reducing the dimension of the graph to different scales. The coarsened subgraph represents the global properties of the original input graph, but at different scales. In the present invention, multiple convolutional layers and pooling layers are used alternately to analyze the low-level and high-level properties of the graph. The number of layers used is a parameter in the design of the neural network.

[0086] In the pooling layer, if the information gain of a vertex is below a threshold, the vertex is removed. The threshold is determined by calculating the average information gain of all vertices in the input graph. The focus of classification is to find the vertices in the training graph whose properties are most useful for distinguishing the class labels. The information gain of the vertex helps determine the vertices that are most relevant for predicting the class label, because the information gain is the information contained in a vertex, and the vertex that can accurately classify the graph into the possible class labels has the largest effective information.

[0087] Given a graph G = (V, E), where v = {v1, v2, ..., v n} is the vertex set of graph G, E is the edge set of graph, Y={y1,y2,……,y S} is a set of class labels. Then the vertex information gain IG(v j ) is expressed as:

[0088]

[0089] Among them, P(y s ) is the label y s The probability, P(y s |v j ) is based on the j-th vertex v j Attribute data classifies graph data into labels y s probability.

[0090] The following Algorithm 3 shows the working principle of the pooling layer in the neural network of the present invention. The goal of the pooling layer is to zv =(V zv ,E) Generate a coarser graph G C =(V C ,E C ),in is the vertex set of the coarsened graph, is the edge set of the coarsened graph. At the same time, the number of vertices and edges in the coarsened graph is less than the number of vertices and edges in the original input graph. That is, |V C |<<|V zv | and|E C|<<|E|. G zv is the input map of the new attribute representation from the convolutional layer, which serves as the input of the pooling layer.

[0091] Algorithm 3

[0092] Input parameters: Input graph G zv =(V zv ,E), Graph data with new attribute representation from graph convolutional layers

[0093] Output parameter: coarsened graph G C =(V C ,E C ), where |V C |<<|V zv |and| E C|<<|E|

[0094] Algorithm parameters: Threshold = average value of information gain (IG) of input graph

[0095] 1.Threshold←IG average value of input graph

[0096] 2. For each vertex v∈V zv Do the following:

[0097] 2.1.1 If IG(v)<Threshold then

[0098] Prune vertex v and all its connected edges

[0099] 2.1.2 Otherwise

[0100] G C ←G z

[0101] 2.1.3 End condition judgment

[0102] 3. End the loop

[0103] 4. Return to G C

[0104] The above algorithm finally outputs G C This is the final output of the pooling layer.

[0105] Step 2.4, after the softmax normalization function, output the predicted probability score;

[0106] Step 2.5, calculate the error based on the predicted probability score and the score corresponding to the actual label;

[0107] Step 2.6, back-propagate the error to update the parameters of the neural network, including the filter vector, the number of filters, and the weight matrix, until the error meets the requirements and the training is completed;

[0108] Back propagation: The present invention receives the input graph and generates the probability scores of the output labels. Then, the following cross entropy error function is used

[0109]

[0110] The error or cost of the classification is calculated. The error is used to update the parameters of the model such as the filter vector, number of filters, and weight matrix through back propagation.

[0111] Step 3: Identify the open-loop and closed-loop states of the high-proportion renewable energy distribution network based on the trained graph convolutional pooling neural network;

[0112] Step 3.1, collect the real-time operation data of the target high-proportion new energy distribution network, and refer to steps 1.1-1.4 to generate weighted attribute graph data;

[0113] Step 3.2, input the weighted attribute graph data into the trained graph convolutional pooling neural network;

[0114] Step 3.3, obtain the predicted probability score given by the graph convolutional pooling neural network;

[0115] Step 3.4, identifying the open-loop and closed-loop states of the high-proportion renewable energy distribution network based on the predicted probability scores;

[0116] Step 4: The operator performs corresponding feedback operations according to the operating specifications based on the open-loop and closed-loop status identification results of the distribution network.

[0117] The above shows and describes the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above specific embodiments. The above specific embodiments and the description in the specification are only for further illustrating the principles and preparation effects of the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, which fall within the scope of the present invention. The scope of the present invention is defined by the claims and their equivalents.

Claims

1. A GCNN-based open-loop and closed-loop state recognition method for a high-proportion renewable energy distribution network, characterized in that: The steps include: Step 1, mapping the node devices of the target high-proportion new energy distribution network and the electrical connection relationship between the node devices into vertex sets and edge sets respectively, constructing graph data including vertex sets, edge sets, weight matrices and vertex attribute matrices, preprocessing the node data of the target high-proportion new energy distribution network, assigning the preprocessed node data of the target high-proportion new energy distribution network to the graph data, and obtaining a labeled weighted attribute graph data set; Step 2, using the labeled weighted attribute graph data set to train a graph convolutional neural network including at least one convolutional layer and a prediction layer, each convolutional layer performs convolution and pooling operations on the input graph data, obtains the graph data after aggregating the attributes of k-order neighbor vertices and the coarsened graph, the prediction layer normalizes the coarsened graph output by the last convolutional layer, obtains the predicted label characterizing the open-loop and closed-loop state of the target high-proportion new energy distribution network, back-propagates the error between the predicted label and the true label, and updates the graph convolutional neural network parameters; Step 3: Collect node data of the target high-proportion new energy distribution network, obtain a weighted attribute graph data set, input the weighted attribute graph data set into the trained graph convolutional neural network, and obtain the recognition result of the open-loop and closed-loop state of the target high-proportion new energy distribution network.

2. According to claim 1, the method for identifying open-loop and closed-loop states of a high-proportion new energy distribution network based on GCNN is characterized in that: In step 1, the vertex attribute matrix includes attribute data of each node device, and the attribute data of each node device includes but is not limited to: specific operation data and text information. The specific operation data includes but is not limited to: temperature and humidity, voltage, and current. The text information includes but is not limited to: normal temperature and high temperature.

3. According to claim 2, the method for identifying open-loop and closed-loop states of a high-proportion new energy distribution network based on GCNN is characterized in that: In step 1, the edge set includes but is not limited to the following information: whether the target high-proportion new energy distribution network node devices are topologically connected and the edge weight of the topological connection between the node devices, and the edge weight of the topological connection between the node devices is represented by the transmission power information between the node devices.

4. According to claim 3, the method for identifying open-loop and closed-loop states of a high-proportion new energy distribution network based on GCNN is characterized in that: In step 2, the k-order neighbor vertices are vertices corresponding to neighbor nodes whose number of nodes separated by the shortest topological connection between the nodes corresponding to each vertex is k.

5. According to claim 4, the method for identifying open-loop and closed-loop states of a high-proportion new energy distribution network based on GCNN is characterized in that: In step 2, obtaining the aggregated k-order neighbor vertex attributes is achieved by averaging the aggregated k-order neighbor vertex attribute vectors. Among them, N k (v i ) is the i-th vertex v i The vertices in the k-order neighborhood of |N k (v i )| is the number of k-order neighbor nodes of the i-th vertex, The vector after the attributes of the k-order neighbor vertices of the i-th vertex are aggregated. is the jth neighbor vertex v j The attribute vector, w ij For the i-th vertex v i and the jth neighbor vertex v j The weight between ij By i and the jth neighbor vertex v j The edge weights of multiple connected topological lines are summed up and obtained.

6. According to claim 5, the method for identifying open-loop and closed-loop states of a high-proportion new energy distribution network based on GCNN is characterized in that: In step 2, the graph data input to the current convolutional layer has a vertex attribute matrix composed of a vector obtained by aggregating the attributes of k-order neighbor vertices in the previous convolutional layer and a vertex attribute matrix of the graph data output by the current convolutional layer.

7. According to claim 6, the method for identifying open-loop and closed-loop states of a high-proportion new energy distribution network based on GCNN is characterized in that: In step 2, the specific method for each convolution layer to perform pooling operation on the input graph data is: according to the information gain of the vertex set of the graph data input to the current convolution layer, the graph data after the current convolution layer aggregates the attributes of the k-order neighbor vertices is downsampled.

8. According to claim 7, the method for identifying open-loop and closed-loop states of a high-proportion new energy distribution network based on GCNN is characterized in that: The specific method for downsampling the graph data after the current convolutional layer aggregates the attributes of k-order neighbor vertices is: traversing the vertex set of the graph data after the current convolutional layer aggregates the attributes of k-order neighbor vertices, and performing pruning operations on vertices whose information gain is less than a threshold.

9. According to claim 8, the method for identifying open-loop and closed-loop states of a high-proportion new energy distribution network based on GCNN is characterized in that: According to the expression Calculate the information gain of the vertex, where IG(v j ) is the jth vertex v j The information gain, P(y s ) is the label y s The probability, P(y s |v j ) is based on the j-th vertex v j Attribute data classifies graph data into labels y s probability.

10. The open-closed loop state recognition system of high-proportion new energy distribution network based on GCNN is characterized by: include: Operation status sensors at each new energy node, power detection devices between each link, and a host computer for data processing and status identification; The host computer receives data collected by the operating status sensor and the power detection device, and executes the method described in any one of claims 1 to 9.

Images