Low-voltage distribution network node voltage prediction method based on total element learning graph neural network

Abstract

The invention provides a low-voltage power distribution network node voltage prediction method based on a total element learning graph neural network, and belongs to the technical field of smart power grids, and the method comprises the following steps: S1, based on the deficiency of a low-voltage power distribution network topological structure, building a graph neural network model of a low-voltage power distribution network through the iterative updating of the edges of a graph neural network; s2, performing iterative learning of the graph neural network model based on the edge features and the node features of the graph neural network model; and S3, realizing direct prediction of the node voltage of the low-voltage distribution network based on the graph neural network model obtained by learning. According to the method, topological structure characteristics and measurement data characteristics can be considered, direct prediction of the node voltage of the low-voltage distribution network is realized, the model complexity is lower, the potential relationship between data is deeply mined, the prediction model establishment efficiency is higher, and the accuracy of the prediction model is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of smart grids, and specifically relates to a method for predicting node voltages in a low-voltage distribution network based on a full-factor learning graph neural network. Background Art

[0002] As the terminal network of the power system, the low-voltage distribution network has a large number of terminal devices, various types, low intelligence, inconsistent standards, and irregular network wiring. These problems bring huge challenges to the operation and maintenance of the low-voltage distribution network. The power supply reliability and power quality often do not meet user requirements, and voltage over-limit is one of the common problems. Therefore, accurate prediction of the node voltage of the low-voltage distribution network can effectively prevent potential power hazards.

[0003] The method for predicting node voltage in distribution network based on power flow calculation mainly calculates node voltage based on Kirchhoff's law on the basis of impedance matrix calculation. However, the traditional power flow calculation method may encounter convergence problems in some cases, that is, the iterative algorithm may not converge to a stable solution, and certain operating conditions of the power system may cause the power flow equation to become an ill-conditioned problem, which will significantly affect the accuracy and stability of the calculation.

[0004] Conventional node voltage prediction methods are mostly proposed for medium and high voltage distribution networks, and only some methods that ignore the topological structure of the distribution network can be applied to low voltage distribution networks. However, research shows that the node voltage prediction method using topological structure information is better than the method without topological information. Therefore, the above low voltage distribution network node voltage prediction method cannot consider the topological structure characteristics of the low voltage distribution network, and the potential relationship between the data is not mined enough, resulting in low model accuracy. Summary of the invention

[0005] In order to solve the above problems existing in the prior art, a low-voltage distribution network node voltage prediction method based on a full-factor learning graph neural network is provided.

[0006] The technical solution adopted by the present invention to solve its technical problem is:

[0007] This technical solution proposes a low-voltage distribution network node voltage prediction method based on a full-factor learning graph neural network, including the following steps:

[0008] S1: Based on the missing topological structure of the low-voltage distribution network, a graph neural network model of the low-voltage distribution network is established by iteratively updating the edges of the graph neural network;

[0009] S2: Iteratively learn the graph neural network model based on the edge features and node features of the graph neural network model;

[0010] S3: Based on the learned graph neural network model, direct prediction of low-voltage distribution network node voltage is achieved.

[0011] Preferably, in S1, the graph neural network model includes edge features, node features and global features.

[0012] Preferably, the low-voltage distribution network topology is regarded as a complete graph, and the edge characteristics include the conductance and admittance of the corresponding line. The weight of the edge with actual connection is the admittance value of the line segment, and the weight of the edge with no connection is 0.

[0013] Preferably, the node characteristics include three electrical quantities of the circuit node, the three electrical quantities are active power, reactive power and voltage, which are set as per-unit values ​​when input.

[0014] Preferably, the global features include: the node voltage of the transformer is used as a global variable, and the vertex is represented as:

[0015] v i = {P i ,Q i ,U i} (1);

[0016] Where P i , Q i , U i Respectively represent the active power, reactive power and voltage of node i, and the characteristic attribute is a three-dimensional vector;

[0017] The edge is represented as:

[0018] e k = {G k ,B k} (2);

[0019] In the formula, G k , B k They represent the conductance and susceptance of edge k respectively, and the characteristic attribute is a two-dimensional vector.

[0020] Preferably, it is characterized in that the graph neural network model with global variables is expressed as:

[0021] G(u0,V,E)(3);

[0022] Where u0 is the global variable representing the transformer voltage, V = {v i} is a vertex set consisting of user nodes and intermediate nodes, E = {e k ,r k ,s k} is the set of edges represented by triples;

[0023] Among them, e kis the feature of edge k, r k is the index of the receiving node, s k The index of the sending node.

[0024] Preferably, the iterative learning process of the graph neural network model includes:

[0025] Edge feature learning: Defining edge feature update function The formula is:

[0026]

[0027] In the formula, e' k is the updated feature of edge k, is the feature of the receiving node of edge k, is the characteristic of the sending node.

[0028] Preferably, It is implemented by a three-layer fully connected network, whose input is a 9-dimensional feature vector, expressed as:

[0029]

[0030] The output is a 2D feature vector representing the conductance and susceptance of the edge [G' k ,B' k ];

[0031] Node feature learning: Defining edge aggregation function ρ e→v , the formula is:

[0032]

[0033] In the formula, The edge features are aggregated into the adjacency matrix, which is converted into a weighted adjacency matrix. It consists of two layers of fully connected networks, and the input is the edge feature [G k ,B k ], the output is 1-dimensional features.

[0034] Preferably, define a node update function The formula is:

[0035]

[0036] The node update function is implemented using a 3-layer graph convolutional neural network;

[0037] Vertex feature matrix H at time l+1 (l+1) Calculated by formula (8), the formula is expressed as:

[0038] H (l+1) =GCN(H (l) ) (8);

[0039] In the formula, H (l) is the vertex feature matrix at the lth moment. GCH is implemented by three layers of graph convolutional layers. The calculation method of each layer is expressed as follows:

[0040]

[0041] Where i = 1, 2, 3, and are the inputs of the i-th layer and the (i-1)-th layer respectively, W is the weight matrix, when i=1, When i=3,

[0042] is the adjacency matrix with self-loops added, for The degree matrix of .

[0043] Preferably, a regularization term about global variables is added to the first layer of GCN, and the formula is expressed as:

[0044]

[0045] In the formula, is the learning weight of the global variable u0, is the weighted adjacency matrix with self-loops added.

[0046] Compared with the prior art, the present invention has the following advantages:

[0047] 1. Under the condition that the topological structure of the low-voltage distribution network is missing, this application can realize the autonomous evolution of the potential topology by iteratively updating the edges of the graph neural network, so that the connection relationship between the nodes is close to the real low-voltage distribution network, thereby taking into account the topological structure characteristics and the metering data characteristics, and realizing the direct prediction of the node voltage of the low-voltage distribution network.

[0048] 2. This application can clearly reflect the information flow and inter-node relationships between low-voltage distribution network nodes, and can make full use of the neighborhood information between nodes, thereby improving the accuracy of the prediction model and deeply mining the potential relationships between data, thereby improving the performance of data analysis. Compared with the flow calculation model, the model complexity is lower and the efficiency of establishing the prediction model is higher. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] The above and / or additional aspects and advantages of the present invention will become apparent and easily understood from the description of the embodiments in conjunction with the following drawings, in which:

[0050] Figure 1 It is the overall flow chart of the present invention;

[0051] Figure 2 It is the graph neural network model structure of low-voltage distribution network. DETAILED DESCRIPTION

[0052] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and cannot be understood as limiting the present invention.

[0053] like Figure 1-Figure 2 As shown, this embodiment proposes a low-voltage distribution network node voltage prediction method based on a full-factor learning graph neural network, including the following steps:

[0054] S1: Based on the missing topological structure of the low-voltage distribution network, a graph neural network model of the low-voltage distribution network is established by iteratively updating the edges of the graph neural network;

[0055] S2: Iteratively learn the graph neural network model based on the edge features and node features of the graph neural network model;

[0056] S3: Based on the learned graph neural network model, direct prediction of low-voltage distribution network node voltage is achieved.

[0057] In the absence of a low-voltage distribution network topology, the autonomous evolution of the potential topology can be achieved by iteratively updating the edges of the graph neural network. The connection relationship between nodes can be inferred and autonomously evolved through measurement data, making the connection relationship between nodes close to the real low-voltage distribution network, thereby taking into account both the topological structure characteristics and the metering data characteristics, and realizing direct prediction of the node voltage of the low-voltage distribution network.

[0058] In S1, the graph neural network model includes edge features, node features and global features.

[0059] The topological structure of the low-voltage distribution network is regarded as a complete graph. The edge features include the conductance and admittance of the corresponding line. The weight of the edge with actual connection is the admittance value of the line segment, and the weight of the edge without connection is 0. The direction of power transmission is mainly determined by the potential difference between the two ends of the circuit. When using graph neural network to model the low-voltage distribution network, an undirected graph is constructed to accurately reflect this physical phenomenon.

[0060] The node characteristics include three electrical quantities of the circuit node, namely active power, reactive power and voltage. Since the voltage is unknown in the initial state, it is set to a per-unit value when input.

[0061] The global features include: the node voltage of the transformer is used as a global variable, and the graph neural network model of the low-voltage distribution network can be described as follows Figure 2 As shown, Figure 2 It is the graph neural network model structure of the low-voltage distribution network, where the solid line represents the actual connection relationship between two adjacent nodes, and the dotted line represents the absence of an actual connection relationship.

[0062] Figure 2 In , the vertices are represented as:

[0063] v i = {P i ,Q i ,U i} (1);

[0064] Where P i , Q i , U i Respectively represent the active power, reactive power and voltage of node i, and the characteristic attribute is a three-dimensional vector;

[0065] The edge is represented as:

[0066] e k = {G k ,B k} (2);

[0067] In the formula, G k , B k They represent the conductance and susceptance of edge k respectively, and the characteristic attribute is a two-dimensional vector.

[0068] The graph neural network model with global variables is represented as:

[0069] G(u0,V,E) (3);

[0070] Where u0 is the global variable representing the transformer voltage, V = {v i} is a vertex set consisting of user nodes and intermediate nodes, E = {e k ,r k ,s k} is the set of edges represented by triples;

[0071] Among them, e k is the feature of edge k, r k is the index of the receiving node, s k The index of the sending node.

[0072] The iterative learning process of the graph neural network model includes:

[0073] Edge feature learning: Defining edge feature update function The formula is:

[0074]

[0075] In the formula, e'k is the updated feature of edge k, is the feature of the receiving node of edge k, is the characteristic of the sending node;

[0076] It is implemented by a three-layer fully connected network, whose input is a 9-dimensional feature vector, expressed as:

[0077]

[0078] The output is a 2D feature vector representing the conductance and susceptance of the edge [G' k ,B' k ];

[0079] Node feature learning: Defining edge aggregation function ρ e→v , the formula is:

[0080]

[0081] In the formula, The edge features are aggregated into the adjacency matrix, which is converted into a weighted adjacency matrix. It consists of two layers of fully connected networks, and the input is the edge feature [G k ,B k ], the output is 1-dimensional features.

[0082] Define node update function The formula is:

[0083]

[0084] The node update function is implemented using a 3-layer graph convolutional neural network;

[0085] Vertex feature matrix H at time l+1 (l+1) Calculated by formula (8), the formula is expressed as:

[0086] H (l+1) =GCN(H (l) ) (8);

[0087] In the formula, H (l) is the vertex feature matrix at the lth moment. GCH is implemented by three layers of graph convolutional layers. The calculation method of each layer is expressed as follows:

[0088]

[0089] Where i = 1, 2, 3, and are the inputs of the i-th layer and the (i-1)-th layer respectively, W is the weight matrix, when i=1, When i=3,

[0090] is the adjacency matrix with self-loops added, for The degree matrix of .

[0091] In order to compensate for the fact that formula (9) cannot process global features, a regularization term about global variables is added to the first layer of GCN. The formula is expressed as:

[0092]

[0093] In the formula, is the learning weight of the global variable u0, is the weighted adjacency matrix with self-loops added.

[0094] The learning process of the graph neural network model for low-voltage distribution network node grid prediction is as follows:

[0095] Input:G(u,V,E)

[0096] Output:G(u',V',E')

[0097] 1:fore k ∈E

[0098] 2: / *Calculate updated edge attributes* /

[0099] 3: end for

[0100] 4:forv i ∈Vdo

[0101] 5: / * Aggregate edge attributes of each node * /

[0102] 6: / *Calculate updated node attributes* /

[0103] 7:end for.

[0104] This application can clearly reflect the information flow and inter-node relationships between low-voltage distribution network nodes, and can make full use of the neighborhood information between nodes, improve the accuracy of the prediction model, and deeply mine the potential relationships between data, thereby improving the performance of data analysis. Compared with the power flow calculation model, the model complexity is lower and the efficiency of establishing the prediction model is higher.

[0105] Although the embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the claims and their equivalents.

Claims

1. A low-voltage distribution network node voltage prediction method based on a full-factor learning graph neural network, characterized in that: The following steps are involved: S1: Based on the missing topological structure of the low-voltage distribution network, a graph neural network model of the low-voltage distribution network is established by iteratively updating the edges of the graph neural network; S2: Iteratively learn the graph neural network model based on the edge features and node features of the graph neural network model; S3: Based on the learned graph neural network model, direct prediction of low-voltage distribution network node voltage is achieved.

2. The method for predicting node voltage in a low-voltage distribution network based on a full-factor learning graph neural network according to claim 1 is characterized in that: In S1, the graph neural network model includes edge features, node features and global features.

3. The method for predicting node voltage in a low-voltage distribution network based on a full-factor learning graph neural network according to claim 2 is characterized in that: The topological structure of the low-voltage distribution network is regarded as a complete graph. The edge characteristics include the conductance and admittance of the corresponding line. The weight of the edge with actual connection is the admittance value of the line segment, and the weight of the edge without connection is 0.

4. The method for predicting node voltage in a low-voltage distribution network based on a full-factor learning graph neural network according to claim 2 is characterized in that: The node characteristics include three electrical quantities of the circuit node, which are active power, reactive power and voltage, and are set as per-unit values ​​when input.

5. The method for predicting node voltage in a low-voltage distribution network based on a full-factor learning graph neural network according to claim 2 is characterized in that: The global features include: the node voltage of the transformer is used as a global variable, and the vertex is represented as: v i ={P i ,Q i ,U i } (1); Where P i , Q i , U i Respectively represent the active power, reactive power and voltage of node i, and the characteristic attribute is a three-dimensional vector; The edges are represented as: e k ={G k ,B k } (2); In the formula, G k , B k They represent the conductance and susceptance of edge k respectively, and the characteristic attribute is a two-dimensional vector.

6. The method for predicting node voltage in a low-voltage distribution network based on a full-factor learning graph neural network according to claim 5 is characterized in that: The graph neural network model with global variables is represented as: G(u0,V,E)(3); Where u0 is the global variable representing the transformer voltage, V = {v i } is a vertex set consisting of user nodes and intermediate nodes, E = {e k ,r k ,s k } is the set of edges represented by triples; Among them, e k is the feature of edge k, r k is the index of the receiving node, s k The index of the sending node.

7. The method for predicting node voltage in a low-voltage distribution network based on a full-factor learning graph neural network according to claim 6 is characterized in that: Iterative learning process of graph neural network model include: Edge feature learning: Defining edge feature update function The formula is: In the formula, e' k is the updated feature of edge k, is the feature of the receiving node of edge k, is the characteristic of the sending node.

8. The method for predicting node voltage in a low-voltage distribution network based on a full-factor learning graph neural network according to claim 7 is characterized in that: It is implemented by a three-layer fully connected network, whose input is a 9-dimensional feature vector, expressed as: The output is a 2D feature vector representing the conductance and susceptance of the edge [G' k ,B' k ]; Node feature learning: Defining edge aggregation function ρ e→v , the formula is: In the formula, The edge features are aggregated into the adjacency matrix, which is converted into a weighted adjacency matrix. It consists of two layers of fully connected networks, and the input is the edge feature [G k ,B k ], the output is 1-dimensional features.

9. The method for predicting node voltage in a low-voltage distribution network based on a full-factor learning graph neural network according to claim 8 is characterized in that: Define node update function The formula is: The node update function is implemented using a 3-layer graph convolutional neural network; Vertex feature matrix H at time l+1 (l+1) Calculated by formula (8), the formula is expressed as: H (l+1) =GCN(H (l) ) (8); In the formula, H (l) is the vertex feature matrix at the lth moment. GCH is implemented by three layers of graph convolutional layers. The calculation method of each layer is expressed as follows: Where i = 1, 2, 3, and are the inputs of the i-th layer and the (i-1)-th layer respectively, W is the weight matrix, when i=1, When i=3, is the adjacency matrix with self-loops added, for The degree matrix of .

10. The method for predicting node voltage in a low-voltage distribution network based on a full-factor learning graph neural network according to claim 9 is characterized in that: Add a regular term about global variables in the first layer of GCN, the formula is expressed as: In the formula, is the learning weight of the global variable u0, is the weighted adjacency matrix with self-loops added.

Images