G3-plc communication topology based on gnn to restore power grid physical topology method

Abstract

The application relates to the technical field of smart grids, in particular to a G3-PLC communication topology-based power distribution network physical topology restoration method based on GNN. The method constructs a dynamic graph model of a power distribution network area physical topology according to G3-PLC communication data, inputs the dynamic graph model into a graph neural network model, and the graph neural network model outputs node level distribution and edge connection probability. The power distribution network area physical topology is dynamically updated according to the node level distribution and the edge connection probability, intelligent prediction and restoration of the power distribution network area physical topology are realized without increasing additional hardware devices and any manual operation, the cost is reduced, and the efficiency and reliability of dynamic updating of the power distribution network area physical topology are remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of smart grid, in particular to a G3-PLC communication topology restoring power distribution network physical topology method based on GNN (Graph Neural Networks). BACKGROUND

[0002] The power distribution network area physical topology refers to the physical connection relationship and hierarchical relationship between various devices (such as transformers, switches, lines, smart meters, etc.) in the power distribution network. Accurate identification of the physical topology of the power distribution network is very important for the stable operation of the power system.

[0003] The current power distribution network physical topology identification is mainly the characteristic current injection method: by installing a characteristic current injection device at each branch and the end of the power distribution network, injecting a characteristic current into the power distribution network, and then sending the characteristic current by each device in turn to identify each branch and level of the power distribution network, for example, the low-voltage power distribution area topology dynamic identification method based on characteristic current disclosed in patent application No. CN114665469A. However, this method can accurately identify the physical topology of the power distribution network, but it needs to install corresponding characteristic current injection devices on the existing power distribution network, resulting in high cost. SUMMARY

[0004] Therefore, it is necessary to provide a G3-PLC communication topology restoring power distribution network physical topology method based on GNN to solve the problem that the existing power distribution network area physical topology identification needs to increase additional hardware devices, resulting in high cost.

[0005] The present application provides a G3-PLC communication topology restoring power distribution network physical topology method based on GNN. The method comprises:

[0006] Step S1, constructing a dynamic graph model of the power distribution network area physical topology according to G3-PLC communication data;

[0007] Step S2, inputting the dynamic graph model into a graph neural network model, the graph neural network model outputs node level distribution and edge connection probability; the graph neural network model comprises an input layer, a spatial feature aggregation layer and a hierarchical-topology joint prediction layer connected in cascade, the input layer performs feature extraction and coding on the dynamic graph model and outputs a coded node feature matrix and a coded edge feature matrix to the spatial feature aggregation layer, the spatial feature aggregation layer extracts node spatial aggregation features from the coded node feature matrix and the coded edge feature matrix using a graph attention network, and the hierarchical-topology joint prediction layer performs node level prediction and edge prediction according to the node spatial aggregation features, and outputs the node level distribution and the edge connection probability;

[0008] Step S3: Dynamically update the physical topology of the distribution network area based on the node hierarchy distribution and the edge connection probability.

[0009] Furthermore, the G3-PLC communication data includes communication node data and communication routing data, and step S1 includes:

[0010] Step S11: Construct the node set of the dynamic graph model based on the communication node data; the communication node data includes the hop count from the node device to the root node, the G3-PLC signal propagation delay from the node device to the root node, the metering sampling voltage value of the node device, the metering power value of the node device, the transmit power of the node device, the received signal strength of the node device, and the signal-to-noise ratio of the node device.

[0011] Step S12: Generate candidate physical connection edges based on the G3-PLC neighbor table, and calculate the edge weights of the candidate physical connection edges according to the communication routing data information. The candidate physical connection edges and the edge weights constitute the edge set of the dynamic graph model. The communication routing data includes G3-PLC signal attenuation, G3-PLC signal delay distance, minimum routing hop count for G3-PLC communication, and link quality indication of G3-PLC communication nodes.

[0012] Furthermore, the formula for calculating the edge weight is:

[0013]

[0014] In the formula, , , , ; This represents the edge weight between node i and node j. This indicates the attenuation of the G3-PLC signal from node i to node j. This represents the G3-PLC signal delay distance from node i to node j. This represents the difference in the minimum route hop count for G3-PLC communication between node i and node j. This represents the difference in link quality indicators between node i and node j. , , and All are weighting coefficients.

[0015] Furthermore, the calculation formula for the link quality indicator is as follows:

[0016] LQI i =(SNR i +10)*4

[0017] In the formula, LQI iLink quality indicator for node i, SNR i Signal-to-noise ratio for node i.

[0018] Further, the elements of the node feature matrix are the transmit power, node type, received signal strength, G3-PLC signal time delay distance, link quality indicator, metering sampling voltage value, metering power value and historical level sequence of each node in T time windows; and the elements of the edge feature matrix are the edge weight and historical level difference of each candidate physical connection edge in T time windows.

[0019] Further, the expression of the training loss function of the graph neural network model is:

[0020]

[0021] In the formula, L total is a training loss value, and ξ, ψ and ζ are weight coefficients, L level is a level prediction loss value, L edge is an edge prediction loss value, L constraint is a PLC topology constraint loss value.

[0022] The expression of the level prediction loss function is:

[0023]

[0024] In the formula, L i denotes the level of the i-th node predicted by the graph neural network model, L i * denotes the real level of the i-th node.

[0025] The expression of the edge prediction loss function is:

[0026]

[0027] In the formula, A uv * is the real edge label of the node pair (u, v), P edge is the edge connection prediction probability of the node pair (u, v).

[0028] The expression of the PLC topology constraint loss function is:

[0029]

[0030] In the formula, I1(.) and I2(.) are both indicator functions, ε is a minimum value, h v is the level of the parent node v, h u is the level of the child node u, V vV is a metering sampling voltage value of a parent node v u V is a metering sampling voltage value of a child node u.

[0031] Further, the formula for calculating the attention coefficient of the spatial feature aggregation layer is:

[0032]

[0033] In the formula, a ij represents the attention coefficient of node j to node i, f i represents the encoded node feature of node i, f j represents the encoded node feature of node j, W is a learnable weight matrix, || represents a vector concatenation operation, a is an attention weight, LeakyReLu(.) is an activation function, and softmax(.) is a normalization function.

[0034] Further, the graph attention network adopts a hierarchical attention mechanism, and the attention weight is the product of the basic weight of a node and a hierarchical factor.

[0035] Further, the edge connection probability satisfies a tree-like hierarchical constraint, a voltage drop constraint of a power distribution network, and a power constraint of the power distribution network.

[0036] Further, the hierarchical-topological joint prediction layer adopts a multi-layer perceptron to predict the node hierarchy.

[0037] The above-mentioned G3-PLC communication topology restoration power distribution network physical topology method, by constructing a dynamic graph model of the power distribution network substation physical topology according to the G3-PLC communication data, inputting the dynamic graph model into a graph neural network model, the graph neural network model outputting a node hierarchy distribution and an edge connection probability, and dynamically updating the power distribution network substation physical topology according to the node hierarchy distribution and the edge connection probability, realizes intelligent prediction and restoration of the power distribution network substation physical topology without increasing additional hardware devices and any manual operation, not only reduces the cost, but also significantly improves the efficiency and reliability of dynamic updating of the power distribution network substation physical topology. BRIEF DESCRIPTION OF DRAWINGS

[0038] Figure 1 FIG. 1 is a flowchart of a G3-PLC communication topology restoration power distribution network physical topology method based on GNN in an embodiment. DETAILED DESCRIPTION

[0039] In order to make the purposes, technical solutions and advantages of the present application clearer, the present application will be further described in detail below with reference to the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and do not limit the present application.

[0040] Embodiment 1

[0041] As shown in the embodiment, a power distribution network station area physical topology identification method is provided, which can be executed by a server and specifically includes the following steps: Figure 1

[0042] Step S1, constructing a dynamic graph model of the power distribution network station area physical topology according to G3-PLC communication data.

[0043] Each node device in the power distribution network station area, such as a transformer (root node), branch box, smart electric energy meter, etc., is deployed with a G3-PLC communication module, and G3-PLC communication data is collected from the G3-PLC communication module of each node device. After obtaining the G3-PLC communication data stream, a continuous time window is divided according to a fixed time length T, the value range of T is a positive integer, and a corresponding dynamic graph model is constructed for the G3-PLC communication data in each time window, and then the dynamic graph models corresponding to each time window are output in time sequence to form a dynamic graph sequence {G t}, t = 1, 2, …, T, which will be input to the graph neural network to realize dynamic mapping from the communication topology to the physical structure. In a preferred embodiment, the size of the time window is fixed, for example, 5 minutes.

[0044] In the embodiment, the G3-PLC communication data includes communication node data and communication routing data. Specifically, step S1 includes:

[0045] Step S11, constructing a node set of the dynamic graph model according to the communication node data.

[0046] The communication node data includes the hop number of the node device i to the root node, i.e., the level h i of the node, the G3-PLC signal propagation delay Δt i of the node device i to the root node, the metering sampling voltage value V i of the node device i (unit: volt V), the metering power value P i of the node device i (unit: watt W), the transmit power P txi of the node device i (unit: dBm), the received signal strength RSSI i of the node device i (unit: dBm), and the signal-to-noise ratio SNR i ​(unit: dB). The communication routing data includes G3-PLC signal attenuation, G3-PLC signal delay distance, G3-PLC communication minimum routing hop count and G3-PLC communication node link quality indication. Wherein, the calculation formula of G3-PLC signal attenuation is shown as formula (1), the calculation formula of G3-PLC signal delay distance is shown as formula (2), the calculation formula of G3-PLC communication minimum routing hop count is shown as formula (3), and the calculation formula of G3-PLC communication node link quality indication is shown as formula (4):

[0047] (1)

[0048] In the formula, represents the G3-PLC signal attenuation from node i to node j, P txij represents the G3-PLC signal transmission power from node i to node j, RSSI ij represents the received signal strength from node i to node j, P txi is the G3-PLC signal transmission power of node i, P txj is the G3-PLC signal transmission power of node j.

[0049] (2)

[0050] In the formula, represents the G3-PLC signal delay distance from node i to node j, represents the G3-PLC signal propagation delay from node i to node j, is the signal transmission speed, and the default is 3x10 8 m / s.

[0051] (3)

[0052] In the formula, represents the G3-PLC communication minimum routing hop count from node i to node j, represents the G3-PLC communication minimum routing hop count from node i to root node r, i.e. the level of node i, represents the G3-PLC communication minimum routing hop count from node j to root node r, i.e. the level of node j.

[0053] LQI i =(SNR i +10)*255 / (53.75+10)=(SNR i +10)*4 (4)

[0054] In the formula, LQI i is the link quality indication of node i, SNR iThe signal-to-noise ratio of the node i.

[0055] Specifically, the node set V includes all devices with G3-PLC communication capability in the power distribution network area, such as transformers, smart meters, branch boxes, etc. deployed with G3-PLC communication modules, and each node is associated with its communication node data, including the node level, G3-PLC signal propagation delay, metering sampling voltage value, metering power value, transmission power, received signal strength and signal-to-noise ratio mentioned in the foregoing.

[0056] Step S12, generating candidate physical connection edges based on the G3-PLC neighbor table, and calculating the edge weights of the candidate physical connection edges according to the communication routing data information, the candidate physical connection edges and the edge weights constitute the edge set of the dynamic graph model.

[0057] In the G3-PLC communication network, the G3-PLC neighbor table of each node records the nodes that can directly communicate with the node. Therefore, by generating candidate physical connection edges based on the G3-PLC neighbor table, all possible physical connections can be covered, and real physical connection edges can be avoided. The edge weight of the candidate physical connection edge integrates the communication characteristics of the node, and the smaller the weight is, the higher the possibility of the physical line connection represented by the edge is. The calculation formula of the edge weight is:

[0058] (5)

[0059] In the formula, , , , ; The edge weight between node i and node j is represented by The G3-PLC signal attenuation from node i to node j is represented by The G3-PLC signal delay distance from node i to node j is represented by The difference between the minimum routing hop count of G3-PLC communication between node i and node j is represented by The difference between the link quality indicators of node i and node j is represented by 、 、 and are weight coefficients. In a preferred embodiment, the sum of each weight coefficient is 1, such as =0.2, =0.2, =0.4, =0.2.

[0060] Step S2, inputting the dynamic graph model into the graph neural network model, and outputting the node level and edge connection probability by the graph neural network model.

[0061] The graph neural network model comprises an input layer, a spatial feature aggregation layer and a hierarchical-topological joint prediction layer connected in cascade. The input layer extracts and encodes features of the dynamic graph model and outputs an encoded node feature matrix and an encoded edge feature matrix to the spatial feature aggregation layer. The input layer specifically comprises a convolution layer, a nonlinear activation function layer and a pooling layer. The convolution layer and the nonlinear activation function layer are used to extract features of the input dynamic graph model to obtain a node feature matrix X e R N×Fx×T and an edge feature matrix E e R M×Fe×T with dimensions of M x Fe x T. In this embodiment, N is the number of nodes, M is the number of edges, Fx is the node feature dimension, and Fe is the edge feature dimension. In this embodiment, Fx = 8, the node features include the transmission power of the node, the node type, the received signal strength, the G3-PLC signal propagation delay, the link quality indicator, the meter sampling voltage value, the meter power value and the historical hierarchical sequence, wherein the node type is used to indicate whether the node is a root node, which can be marked by binary, and the historical hierarchical sequence of the node includes the hierarchical values of the node in the past T time windows, for example, the historical hierarchical sequence of node i can be [5, 4, 3, 4, 4]; Fe = 5, the edge features include the edge weight and the historical hierarchical difference (i.e. the difference between the hierarchical value of the parent node and the hierarchical value of the child node, which is fixed at -1 to strengthen the tree constraint). Therefore, in this embodiment, the elements of the node feature matrix are the transmission power, the node type, the received signal strength, the G3-PLC signal delay distance, the link quality indicator, the meter sampling voltage value, the meter power value and the historical hierarchical sequence of each node in T time windows. The elements of the edge feature matrix are the edge weight and the historical hierarchical difference of each candidate physical connection edge in T time windows. Then, the pooling layer encodes the node feature matrix and the edge feature matrix, i.e. reduces the feature dimension in the time dimension, to obtain an encoded node feature matrix X' e R N×Fx and an encoded edge feature matrix E' e R M×Fe . In this embodiment, the nonlinear activation function layer adopts a ReLU activation function.

[0062] The spatial feature aggregation layer is used to capture the spatial dependence of the tree structure, and specifically adopts a GAT (Graph Attention Network) to extract node spatial aggregation features from the encoded node feature matrix and the encoded edge feature matrix. The GAT introduces an attention mechanism on the basis of a graph neural network (GNN), learns different weights for different neighbors through the attention mechanism, and can improve the accuracy of feature extraction. Specifically, the expression of the GAT is shown in equation (6):

[0063] (6)

[0064] In the equation, f iinput features of the spatial feature aggregation layer, i.e., the encoding node feature matrix or the encoding edge feature matrix, f i ’ output features of the spatial feature aggregation layer, i.e., the node spatial aggregation feature, a ij is the attention coefficient of node j to node i, w ij is the parameter of the convolution kernel, and σ represents a nonlinear activation function. Further, the formula for calculating the attention coefficient of the spatial feature aggregation layer is as follows:

[0065] (7)

[0066] wherein a ij is the attention coefficient of node j to node i, f i is the encoded node feature of node i, f j is the encoded node feature of node j, W is a learnable weight matrix, || represents a vector concatenation operation, a is an attention weight, LeakyReLu(.) is an activation function, and softmax(.) is a normalization function.

[0067] In a preferred embodiment, the graph attention network adopts a hierarchical attention mechanism, and the calculation of the attention weight thereof depends on both the node feature similarity and a preset hierarchical factor. Specifically, the attention weight in formula (7) is the product of a basic weight and the preset hierarchical factor, and the basic weight is obtained according to the node feature similarity. The hierarchical attention mechanism can strengthen the hierarchical constraint of the tree topology through the preset hierarchical factor, so that the feature propagation of the root node can dominate in the topology, thereby further improving the accuracy of the physical topology restoration. Preferably, the value of the preset hierarchical factor is 1.5.

[0068] The hierarchical-topology joint prediction layer is used for simultaneously predicting the node hierarchy and the tree edge relationship. When predicting the node hierarchy, the hierarchical-topology joint prediction layer performs node hierarchy prediction and edge prediction according to the node spatial aggregation feature, and outputs the node hierarchy distribution and the edge connection probability.

[0069] Specifically, the hierarchical-topology joint prediction layer adopts a multilayer perceptron (MLP) to predict the node hierarchy L iZ+. The MLP is a kind of feedforward neural network with at least one hidden layer and capable of nonlinear transformation, and its basic structure comprises an input layer, a hidden layer and an output layer, and the nodes of each layer are connected through weights and biases, and an activation function is used for nonlinear transformation. In the embodiment, the activation function of the MLP is a Softmax function, which is used to realize hierarchical classification; the output of the MLP is a node hierarchical distribution p (Li=k) e [0,1], k is a possible hierarchical level, and the node hierarchical distribution is the probability of the hierarchical level of the node i being k, and k=0 represents a root node, and k>=1 represents the hierarchical level of a child node.

[0070] When the hierarchical-topology joint prediction layer performs edge prediction, first, the inner product of the feature vectors of all possible node pairs (u, v) is calculated according to the node spatial aggregation feature, then the PLC network topology constraint is verified, and the edge connection probability corresponding to the node pair that does not meet the PLC network topology constraint is set to 0, and finally the edge connection probability prediction matrix P is obtained. edge e [0,1] N Wherein, u is a parent node, and v is a child node of u. The PLC network topology constraint includes a tree hierarchical constraint as shown in formula (8), a voltage step-down constraint of a power distribution network as shown in formula (9)-(10) and a power constraint of a power distribution network as shown in formula (11):

[0071] (8)

[0072] In the formula, E T is an edge set of a dynamic graph model, h j is the minimum routing hop count of a child node j to a root node, and h i is the minimum routing hop count of a parent node i to a root node. Since the physical topology of a power distribution network area is a tree structure, the generated dynamic topology graph should be a tree structure, and the hierarchical level of a child node should be one more than that of a parent node.

[0073] (9)

[0074] (10)

[0075] In the formula, V i represents a metering sampling voltage value of a node i, V j represents a metering sampling voltage value of a node j, V min represents a preset minimum metering sampling voltage, V max represents a preset maximum metering sampling voltage, and ΔV max represents a preset maximum metering sampling voltage difference.

[0076] (11)

[0077] In the formula, ∑P in represents the metering power input value of the power distribution network, which is calculated from the metering power input value of each node, ∑P out represents the metering power output value of the power distribution network, which is calculated from the metering power output value of each node, P loss represents the metering power loss value of the power distribution network.

[0078] Further, for the training process of the graph neural network model, the embodiment also designs a training loss function as shown in formula (12):

[0079] (12)

[0080] In the formula, L total is the training loss value, and ξ, ψ and ζ are weight coefficients, L level is the hierarchical prediction loss value, L edge is the edge prediction loss value, and L constraint is the PLC topology constraint loss value.

[0081] The expression of the hierarchical prediction loss function is:

[0082] (13)

[0083] In the formula, L i represents the hierarchical level of the i-th node predicted by the graph neural network model, L i * represents the true hierarchical level of the i-th node.

[0084] The expression of the edge prediction loss function is:

[0085] (14)

[0086] In the formula, A uv * is the true edge label of the node pair (u, v), and is in the form of an adjacency matrix, 1 represents that the edge exists, and 0 represents that the edge does not exist, P edge is the edge connection prediction probability of the node pair (u, v).

[0087] For each node, the tree hierarchical constraint and the voltage step-down constraint must be met, and an additional loss is added for the node that does not meet the two constraint conditions. Specifically, the expression of the PLC topology constraint loss function is:

[0088] (15)

[0089] wherein, I1(.) and I2(.) are both indicator functions, the value of the indicator function is 1 when the condition recorded in the indicator function is true, otherwise 0; ε is a minimum value, h v is the level of the parent node v, h u is the level of the child node u, V v is the metering sampling voltage value of the parent node v, V u is the metering sampling voltage value of the child node u.

[0090] Step S3, dynamically updating the physical topology of the power distribution network according to the node level and the edge connection probability.

[0091] Specifically, when the edge connection probability between node i and node j is greater than or equal to a preset probability threshold, it is determined that there is an edge between node i and node j, and the corresponding edge is added to the edge set of the dynamic graph model; when the edge connection probability between node i and node j is less than the preset probability threshold, it is determined that there is no edge between node i and node j, and the corresponding edge is deleted from the edge set of the dynamic graph model. In addition, when there is a node device to be added or deleted, a corresponding adding or deleting operation is performed on the node set of the dynamic graph model, and then the physical topology is updated according to the updated node set.

[0092] The G3-PLC communication topology restoration method based on GNN of the embodiment, by constructing a dynamic graph model of the physical topology of the power distribution network according to the G3-PLC communication data, inputting the dynamic graph model into the graph neural network model, the graph neural network model outputs the node level distribution and the edge connection probability, and dynamically updating the physical topology of the power distribution network according to the node level distribution and the edge connection probability, realizes intelligent prediction and restoration of the physical topology of the power distribution network without increasing additional hardware devices and any manual operation, not only reduces the cost, but also significantly improves the efficiency and reliability of the dynamic update of the physical topology of the power distribution network.

[0093] The technical features of the above-described embodiments can be combined in any manner. To make the description concise, not all possible combinations of the technical features in the above-described embodiments are described, but as long as the combinations of the technical features do not exist contradictions, they should be considered as the scope of the present application.

[0094] The above-described embodiments only express several implementation manners of the present application, and the description is more specific and detailed, but it should not be understood as a limitation on the scope of the patent. It should be pointed out that for ordinary skilled in the art, without departing from the concept of the present application, a number of modifications and improvements can be made, which are all within the scope of the present application. Therefore, the protection scope of the patent of the present application should be subject to the appended claims.

Claims

1. A method for restoring a physical topology of a power distribution network based on a GNN-based G3-PLC communication topology, characterized in that, The method comprises: Step S1, constructing a dynamic graph model of a power distribution network station area physical topology according to G3-PLC communication data; Step S2, inputting the dynamic graph model into a graph neural network model, and the graph neural network model outputs node level distribution and edge connection probability; the graph neural network model comprises an input layer, a spatial feature aggregation layer and a level-topology joint prediction layer connected in cascade, the input layer performs feature extraction and coding on the dynamic graph model, and outputs a coded node feature matrix and a coded edge feature matrix to the spatial feature aggregation layer, the spatial feature aggregation layer extracts node spatial aggregation features from the coded node feature matrix and the coded edge feature matrix by using a graph attention network, and the level-topology joint prediction layer performs node level prediction and edge prediction according to the node spatial aggregation features, and outputs the node level distribution and the edge connection probability; Step S3, dynamically updating the power distribution network station area physical topology according to the node level distribution and the edge connection probability.

2. The method of claim 1, wherein the GNN-based G3-PLC communication topology restoration power distribution network physical topology method further comprises: The G3-PLC communication data comprises communication node data and communication routing data, and the step S1 comprises: Step S11, constructing a node set of the dynamic graph model according to the communication node data; the communication node data comprises the number of hops from a node device to a root node, the G3-PLC signal propagation delay from the node device to the root node, the metering sampling voltage value of the node device, the metering power value of the node device, the transmission power of the node device, the received signal strength of the node device and the signal-to-noise ratio of the node device; Step S12, generating a candidate physical connection edge based on a G3-PLC neighbor table, and calculating an edge weight of the candidate physical connection edge according to the communication routing data information, wherein the candidate physical connection edge and the edge weight constitute an edge set of the dynamic graph model; the communication routing data comprises G3-PLC signal attenuation, G3-PLC signal time delay distance, G3-PLC communication minimum routing hop number and link quality indication of a G3-PLC communication node.

3. The method of claim 2, wherein the GNN-based G3-PLC communication topology restoration power grid physical topology method further comprises: The calculation formula of the edge weight is: ； wherein, , , , ; denotes the edge weight between node i and node j, denotes the G3-PLC signal attenuation from node i to node j, denotes the G3-PLC signal latency distance from node i to node j, denotes the difference of the minimum routing hop count of G3-PLC communication between node i and node j, denotes the difference of the link quality indication between node i and node j; , , and are weight coefficients.

4. The method of claim 2, wherein the GNN-based G3-PLC communication topology restoration power distribution network physical topology method further comprises: The calculation formula of the link quality indication is: LQI i = (SNR i + 10) * 4 In the formula, LQI i is the link quality indicator of node i, SNR i is the signal-to-noise ratio of node i.

5. The method of claim 2, wherein the GNN-based G3-PLC communication topology restoration power grid physical topology method further comprises: The elements of the node feature matrix are the transmission power, node type, received signal strength, G3-PLC signal time delay distance, link quality indication, metering sampling voltage value, metering power value and historical level sequence of each node in T time windows; and the elements of the edge feature matrix are the edge weight and historical level difference of each candidate physical connection edge in T time windows.

6. The method of claim 1, wherein the GNN-based G3-PLC communication topology restoration power grid physical topology method further comprises: The expression of the training loss function of the graph neural network model is: ； In the formula, L total is a training loss value, ξ, ψ and ζ are weight coefficients, L level is a hierarchical prediction loss value, L edge is an edge prediction loss value, L constraint is a PLC topology constraint loss value; The expression of the level prediction loss function is: ； In the formula, L i denotes the predicted level of the i-th node by the graph neural network model, L i * denotes the true level of the i-th node, denotes the probability that the predicted level is the true level. The expression of the edge prediction loss function is: ； In the formula, A uv * is the true edge label of the node pair (u, v), P edge [u][v] is the edge connection prediction probability of the node pair (u, v); The expression of the PLC topology constraint loss function is: ； where I1(.) and I2(.) are indicator functions, ε is a minimum value, h v is the level of the parent node v, h u is the level of the child node u, V v is the metrology sampled voltage value of the parent node v, V u is the metrology sampled voltage value of the child node u, is the maximum voltage difference.

7. The method of claim 1, wherein the GNN-based G3-PLC communication topology restoration power grid physical topology method further comprises: The formula for calculating the attention coefficient of the spatial feature aggregation layer is: ； In the formula, α ij represents the attention coefficient of node j to node i, f i represents the encoded node feature of node i, f j represents the encoded node feature of node j, W is a learnable weight matrix, || represents a vector splicing operation, a is an attention weight, LeakyReLu(.) is an activation function, and softmax(.) is a normalization function.

8. The method of claim 7, wherein the GNN-based G3-PLC communication topology restoration power grid physical topology method further comprises: The graph attention network adopts a level attention mechanism, and the attention weight is the product of the basic weight of a node and a level factor.

9. The method of claim 1, wherein the GNN-based G3-PLC communication topology restoration power distribution network physical topology is further characterized by, The edge connection probability satisfies a tree level constraint, a voltage drop constraint of a power distribution network and a power constraint of the power distribution network.

10. The method of claim 1, wherein the GNN-based G3-PLC communication topology restoration power grid physical topology method further comprises: The level-topology joint prediction layer predicts the node level by using a multi-layer perception machine.

Images