Electric power system target selection method under incomplete information condition

Abstract

The invention discloses a power system target selection method under the condition of incomplete information, and the method comprises the steps: building a topological model of a power target system grid structure, and determining a node feature matrix, a line feature matrix and an adjacent matrix; according to a graph completion learning algorithm based on a multi-level comparison graph automatic coding and decoding network, in a structure channel, a missing graph structure is rebuilt through graph structure diffusion, on a feature channel, missing node features are preliminarily completed through a layer of trainable parameter matrix, meanwhile, node embedding expressions of the structure channel and the feature channel are calculated, and the structure channel and the feature channel are subjected to graph structure reconstruction. After weighted fusion, a classification result is obtained through a classifier, line features are complemented, and a graph completion learning model oriented to the power system is formed; calculating a structure completion confidence coefficient, a node feature completion confidence coefficient, a line feature completion confidence coefficient and a comprehensive confidence coefficient; and calculating the target importance degree of the power system nodes to complete target selection. The objective importance evaluation problem of the power system is solved.

Description

Technical Field

[0001] This invention relates to target selection technology, and more specifically to a target selection method for power systems under conditions of incomplete information. Background Technology

[0002] Accurate identification of key nodes in a power system is a prerequisite for effectively reducing its system efficiency. However, under asymmetric information game conditions, the acquired power system data exhibits fragmentation, heterogeneity, and temporal discretization. Current power system value assessment models based on complete information, when applied in scenarios with missing information, lead to large errors in key node identification. To improve the adaptability of power system value assessment models in scenarios with missing information, graph neural networks can be used to achieve topological completion of missing data. However, this method does not fully evaluate the reliability of the completed information, resulting in unreliable ranking results. Summary of the Invention

[0003] The purpose of this invention is to provide a method for selecting power system objectives under conditions of incomplete information, so as to solve the problem of assessing the importance of power system objectives.

[0004] The technical solution to achieve the objective of this invention is: a method for selecting targets in a power system under conditions of incomplete information, comprising the following steps:

[0005] Step 1: Obtain target-related data of the power system as a training set, which includes power plant node data, substation node data and transmission line data;

[0006] Step 2: Construct a set of nodes using power plants and substations, and a set of edges using transmission line data. Based on this, establish a topological model of the power target system network structure, model the power system as a graph structure, and then determine the node feature matrix, line feature matrix, and adjacency matrix.

[0007] Step 3: Based on the graph completion learning algorithm of the multi-level contrast graph automatic encoding and decoding network, in the structural channel, the missing graph structure is reconstructed by graph structure diffusion. In the feature channel, the missing node features are initially completed by a trainable parameter matrix. At the same time, the node embedding representations of the structural channel and the feature channel are calculated, weighted and fused, and then passed through a classifier to obtain the classification result. The line features are then completed to form a graph completion learning model for power systems.

[0008] Step 4: For the target power system, perform structural completion, node feature completion, and line feature completion based on the trained graph completion learning model for power systems, and calculate the confidence scores of structural completion, node feature completion, line feature completion, and overall confidence score.

[0009] Step 5: Calculate the target importance of power system nodes to support power system target selection under conditions of incomplete information.

[0010] Further, step 1: Obtain target-related data of the power system as a training set, wherein the target-related data of the power system includes power plant node data, substation node data, and transmission line data, wherein:

[0011] The power plant node data includes type code, maximum generating capacity, geographic coordinates, and voltage level. The type code values ​​and corresponding types are 0-'thermal power', 1-'hydro power', 2-'wind power', 3-'solar power', and 4-'nuclear power'.

[0012] Substation node data includes type code, capacity, geographic coordinates, voltage level, transformer model code, turns ratio coefficient, and number of incoming and outgoing line circuits. The type code value is 5.

[0013] Transmission line data includes a connection matrix, as well as voltage level, conductor type, conductor starting station, number of parallel circuits, and total regional load.

[0014] Further, step 2: Construct a node set using power plants and substations, and an edge set using transmission line data. Based on this, establish a topological model of the power target system network structure, modeling the power system as a graph structure, and then determine the node feature matrix, line feature matrix, and adjacency matrix. The specific method is as follows:

[0015] Step 2.1: Construct a set of nodes using power plants and substations, and a set of edges using transmission line data. Based on this, establish a topological model of the power target system network structure, modeling the power system as a graph structure, represented as:

[0016] G train =(V train E train )

[0017] Where V train ={v1,v2,…,v N} represents the set of nodes, including power plants and substations; the set of edges E train ={e ij} represents a power transmission line;

[0018] Step 2.2, determine the node feature matrix, the line feature matrix, and the adjacency matrix, where:

[0019] Node feature matrix X train ∈R N×d N represents the number of nodes, and d represents the feature dimension;

[0020] Line feature matrix F train ∈RM×6 Each line includes: F train = [Voltage level, conductor type, conductor start point, conductor end point, number of parallel circuits, area load], where M is the number of transmission lines;

[0021] Adjacency matrix A train ∈{0,1} N×N If node v i With v j If a power transmission line exists, then A ij =1, otherwise 0.

[0022] Further, in step 3, the graph completion learning algorithm based on a multi-level contrast graph automatic encoding and decoding network reconstructs the missing graph structure in the structural channel through graph structure diffusion. In the feature channel, it initially completes the missing node features through a trainable parameter matrix. Simultaneously, it calculates the node embedding representations of the structural and feature channels, weights and fuses them, and then passes them through a classifier to obtain the classification result. Finally, it performs a full analysis of the line features, forming a graph completion learning model for power systems. The specific method is as follows:

[0023] Step 3.1: Input the structural adjacency matrix A train and node feature matrix X train The graphs are fed into the structure channel and the feature channel, respectively. In the structure channel, the missing graph structure is reconstructed through graph structure diffusion. The reconstructed graph structure is represented as the completed structure adjacency matrix. In the feature channels, a trainable parameter matrix is ​​used to initially complete the missing node features. The reconstructed node feature matrix is ​​represented as follows:

[0024] Step 3.1.1: Input the structural adjacency matrix A train A personalized PageRank algorithm is used to analyze the structural adjacency matrix A. train Perform structural diffusion and set physical constraints on the power system structure during the diffusion process, including prohibiting direct connections between power plant nodes and allowing connections only between nodes of the same voltage level, to generate a completed structural adjacency matrix. in:

[0025] Structural diffusion formula:

[0026]

[0027] in Let A be the structural adjacency matrix at time t. train α is the restart probability, I n It is an identity matrix with the same dimensions as the structural adjacency matrix, initially

[0028] A is removed using k-nearest neighbor sparsification technique SR For unimportant edges, generate the completed structural adjacency matrix. Its elements for:

[0029]

[0030] Where kNN(i) represents the set of the top (k) neighbors of the i-th node, if node v j Belongs to node v i The top-k neighbor set is then retained. Otherwise, set to zero;

[0031] Step 3.1.2: Input the node feature matrix X train ∈R N×d Perform node feature matrix initialization and completion; construct the relationship with X. train Parameter matrices Θ of identical size ij ∈R N×d N is the number of nodes used to fill in missing values ​​in the node feature matrix. Let the initialized and completed node feature matrix be... This is represented as:

[0032]

[0033] Where X i,j For node v i The j-th dimension feature, when X i,j Use random parameter Θ when missing i,j The values ​​are used to fill the values, and physical constraints are applied to the node characteristics during the filling process, including parameters related to the shielded transformer of the power plant node and the fixed type parameters of the substation node.

[0034] Step 3.2: Complete the structural adjacency matrix. The data is fed into a structural encoder, and positional encoding is used as the initial feature to compute the node embedding representation Z of the structural channel. SR ; Complete the structural adjacency matrix The adjacency matrix for guided feature diffusion is obtained by weighted summation. The node features are fed into the feature channel to guide feature diffusion. The diffused node features are then fed into the feature encoder to compute the node embedding representation Z of the feature channel. FR ;

[0035] Step 3.2.1: Construct a structure encoder based on a graph convolutional network, taking the completed structure adjacency matrix as input. The positional encoding (PE) is used, and the structure encoding is performed based on a two-layer graph convolutional neural network (GCN). The output is Z. SR ;

[0036] For the set of nodes Vtrain The location information in the data is modeled, and a location code PE∈R is generated for each node. N×d ,

[0037]

[0038] Where PE(v) i ) for a single node v i Position encoding vector, W∈R N×d It is a learnable weight matrix, b∈R N×d For learnable bias vectors, It is node v i The one-hot encoding is an N-dimensional vector;

[0039] By using a two-layer graph convolutional neural network (GCN), the encoder output of the structural channel is obtained as the node embedding representation Z of the structural channel. SR ;

[0040]

[0041] in This represents the position encoding vector for each node, where σ is a non-linear activation function. This represents the weight matrix of the first and second graph convolutional layers;

[0042] Step 3.2.2: Construct a feature channel encoder based on feature diffusion and multilayer perceptron, and use the completed structural adjacency matrix. The adjacency matrix for guided feature diffusion is obtained by weighted summation. The data is fed into the feature channel to guide feature diffusion, and the diffused node features are then... The data is fed into a feature encoder, and a multilayer perceptron (MLP) is used to compute the feature channel node embedding representation Z. FR ;

[0043]

[0044] Where Λ∈R N×N It is a learnable coefficient matrix. To guide the adjacency matrix of feature diffusion, by It is obtained by weighted multiplication with the coefficient matrix;

[0045] Input the completed node feature matrix Perform node feature diffusion;

[0046]

[0047] Where X (0) X represents the node feature matrix at the initial time step. (t)The matrix represents the node feature matrix at time t, where T represents the number of rounds of diffusion propagation. This represents the node feature matrix output after T rounds of diffusion propagation, where α∈(0,1] is the restart probability;

[0048] A two-layer MLP feature encoder module is introduced to perform message transformation, converting the node feature matrix after diffusion propagation. The data is fed into an MLP feature encoder to obtain the feature channel node embedding representation Z. FR The message conversion process is as follows:

[0049]

[0050] Where W1 and W2 are the learnable weight matrices of the first and second layers of the MLP, respectively, b1 and b2 are the corresponding bias matrices, σ is the nonlinear activation function (the same as the nonlinear activation function in step 3.2.1), and Z... FR H is the node embedding representation of the learned feature channels, where H is the intermediate representation of the node features in the feature channels after the first layer transformation.

[0051] Step 3.3: Input the node embedding representation Z of the structural channel from Step 3.2.1 SR The node embedding representation Z of the feature channels in step 3.2.2 FR Using attention mechanism to fuse Z FR and Z SR The weighted fused node embedding representation Z is then processed by a softmax classifier to obtain the classification result.

[0052] Step 3.3.1: Calculate the node embedding representation Z of the feature channels FR Feature channel weights γ f The node embedding representation Z of the computational structure channel. SR Structure channel weight γ s ;

[0053]

[0054] in and Represents the linear transformation MLP, and W represents the parameters of the MLP. F and W′ F

[0055] Let γ be the learnable weight matrix, tanh be the nonlinear activation function for calculating the weights of the feature and structural channels, and γ be the weight matrix. s and γ fThese represent the raw scores of the feature channel and the structure channel, respectively, calculated initially through the attention mechanism;

[0056] Step 3.3.2: Adjust the feature channel weights γ f and structural channel weight γ s After normalization, the weight coefficients of the two channels are:

[0057]

[0058] Where w s and w f These represent the weights of the feature channel and the structure channel after normalization, respectively.

[0059] Step 3.3.3: Calculate the weighted fused node embedding representation Z;

[0060] Z = w f ×Z FR +w s ×Z SR

[0061] Step 3.3.4: After obtaining the weighted fusion node embedding representation Z in step 3.3.3, use a fully connected layer to reduce the dimension of Z to the category dimension, thereby obtaining the score S of the sample node belonging to each category;

[0062] S = fc(Z)

[0063] Where S∈R n×c c represents the number of categories, s i,j Let fc represent the score of the i-th node belonging to the j-th class, and fc represent the fully connected layer. The softmax function is used to map the score to the (0,1) interval, and make the sum of the probabilities of the node belonging to each class equal to 1.

[0064]

[0065] Finally, the category prediction matrix of the nodes is obtained. in This represents the probability that the i-th node belongs to the j-th class;

[0066] Step 3.4: Analyze the line characteristic F train ∈R M×6 To complete the information, the line characteristics include six features: voltage level, conductor type, conductor start and end points, number of parallel circuits, and area load.

[0067] Step 3.4.1: Perform voltage level completion for each completed line e ij According to the voltage levels V at both ends i and V j Determine the line voltage level:

[0068] V e =max(V i V j )

[0069] Where V e For line e ij The voltage level is determined by taking the higher voltage value at both ends as the line voltage level. If the node voltage is unknown, it is set to a preset value to ensure that the voltage requirement is met. e >35KV;

[0070] Step 3.4.2: Complete the wire type based on voltage level V. e Match the preset model library;

[0071]

[0072] Where X e For the wire type, P(m|V) e ) indicates a known voltage level V e The probability of occurrence of type m, V e =500KV model is LGJ-400, V e =300KV model LGJ-300, V e =250KV model LGJ-240, V e = 110KV model LGJ-150;

[0073] Step 3.4.3: Complete the start and end points of the traverse, and then use the completed adjacency matrix. Extract physical connections and determine line endpoints;

[0074] like Then a line e is generated. ij If node v i For power plants (node ​​v) i If the type encoding is 0, 1, 2, 3, 4, then v i As the starting point of the route S e v j The endpoint is v; i , and v j All are substations (node ​​v) i The type code is 5). Compare the sum of the longitude and latitude coordinates of two nodes, and select the node with the smaller sum as the starting point. Its geographical sorting logic is East → West (the larger the longitude, the further east it is, and the smaller the longitude, the further west it is) and North → South (the larger the latitude, the further north it is, and the smaller the latitude, the further south it is) to determine the starting point.

[0075]

[0076] Step 3.4.4: Complete the parallel circuit count based on the total regional load L. region Predict the number of parallel circuits P e ;

[0077]

[0078] Where β P For the regression coefficients, ensure P e ≥1, L region The total load of the region is defined as clip(x,a,b), which limits x to [a,b]. Step 3.4.5: Perform regional load supplementation, and supplement the current regional load by weighted average of the loads of adjacent regions.

[0079]

[0080] Where L e Let N(e) be the area load of line e, and N(e) be the set of lines adjacent to line e, with weights... dist(e,k) is the geographical distance between lines e and k, ∈ is a minimal constant, and L k The known load value of the adjacent line k.

[0081] Furthermore, a graph completion learning model for power systems is trained using the cross-entropy loss function. The specific method is as follows:

[0082] Step 3.5: Embed the node representation Z of the feature channels FR The data is fed into the feature decoder, and the node channels are used to reconstruct the features. Embedding the nodes of the structural channel to represent Z SR The data is fed into the structure decoder to calculate the structural channel reconstruction features.

[0083] This is used for subsequent calculation of contrastive loss for alignment. and

[0084] Step 3.5.1: Input the node embedding representation Z of the structural channel from Step 3.2.1 SR The node embedding representation Z of the feature channels in step 3.2.2 FR , for Z FR and Z SR Perform random masking;

[0085]

[0086] Among them, MASK z Let represent a mask matrix, where the elements mask∈{0,1}. Z is the node embedding representation of the feature channels. FR The result after masking. It is a structural channel node embedding representation Z SR The result after masking, MASK z With Z FR and Z SR They have the same dimensions;

[0087] Step 3.5.2: Use a single MLP layer to implement decoding, converting Z... FR Input to the feature decoder to compute the reconstructed features of the feature channels. Z SR Input to the structure decoder to compute the reconstructed features of the structure channels.

[0088]

[0089] Where ψ f MLP decoder for feature channels Parameter, ψ s MLP decoder for structured channels parameter.

[0090] Step 3.6: Calculate the classification results Cross-entropy loss with the real data label Y Node embedding representation Z of feature channels FR The node embedding representation of the structural channel Z SR encoder contrast loss Reconstruction features of feature channels and the reconstruction features of structural channels Decoder contrast loss And the total loss is obtained by weighted summation.

[0091] Step 3.6.1: Calculate the classification results Cross-entropy loss with the real data label Y

[0092]

[0093] Where Y i This represents the true label of the i-th sample. Let represent the predicted label of the i-th sample, and n represent the number of labels;

[0094] Step 3.6.2: Calculate the encoder alignment loss during the encoding stage. Extract supervision signals from unlabeled samples to avoid overfitting;

[0095]

[0096] in Let τ be the prototype vector of category j in the feature channel and the structure channel, τ be an adjustable scaling factor, and C be the total number of categories;

[0097]

[0098] Where μ j Let be the prototype vector of category j, where the superscript indicates the feature channel FR or the structural channel SR. The predicted categories for node i are represented by c, where c is the total number of categories, and q is the number of categories. i Represents node v i The weights, Q i The weight x for all nodes predicted as class j i the sum of Filter all nodes v with predicted category j i , For node v of the structural channel in step 3.2.1 i Embedded representation Z SR , For the node v of the feature channel in step 3.2.2 i Embedded representation Z FR ;

[0099] Step 3.6.3: Compare the loss using the MLP decoder for the feature channel and the MLP decoder for the structure channel as described in Step 3.5.2. Supervise the reconstruction process and use contrastive loss alignment. and Calculate the contrast loss of the decoder

[0100]

[0101] in For decoder reconstruction loss, The reconstructed features of the feature channels and structural channels in step 3.5.2 are denoted as , where is the vector inner product, n is the total number of nodes, and τ is the adjustment coefficient.

[0102] Step 3.6.4: Calculate the total loss function using weighted summation.

[0103]

[0104] Where λ1 is the encoder loss The weighting coefficients, where λ² is the decoder loss. Weighting coefficients;

[0105] Step 3.6: Update the parameters of the graph completion learning model for the power system using the gradient backpropagation algorithm until convergence.

[0106] Further, in step 4, for the target power system, structural completion, node feature completion, and line feature completion are performed based on the trained graph completion learning model for power systems. The confidence scores for structural completion, feature completion, line completion, and overall completion are calculated, where:

[0107] (1) Confidence of structural completion

[0108] The confidence level of structural completion reflects the reliability of topology completion, and the computation node v i

[0109] The average connection probability with all other nodes; a higher value indicates more reliable topology completion.

[0110]

[0111] Where A comp ∈R N×N This is the completed probabilistic adjacency matrix. Represents node v i and v j The connection probability, where N is the total number of nodes;

[0112] The node embedding representation Z of the structural channel output by the structural encoder (GCN) in step 3.2.1 SR Generate the completed node connection probabilities using MLP:

[0113]

[0114] Where θ is the Sigmoid function, and its output value is... This represents the probability of node connection.

[0115] (2) Confidence of node feature completion

[0116] The confidence score of node feature completion is used to evaluate the quality of node feature repair. Based on the absolute error between the known node features and the completed node features, it is normalized and then mapped by tanh. Finally, it is converted into the confidence score of node feature completion by 1-tanh(·).

[0117]

[0118] in X is the completed node feature matrix. known Given the known node feature values, M∈{0,1} N×D The missing mask matrix is ​​denoted by ∈, where ∈ is the smoothing coefficient and D is the number of feature dimensions.

[0119] (3) Confidence of line feature completion

[0120] The confidence level of line feature completion is calculated by weighting the connection probability of the two-end nodes and the verification results of physical constraints.

[0121]

[0122] in Let S be the node connection probability of nodes s and t at both ends of line e in the completed adjacency matrix. For voltage effectiveness, For the validity of the number of parallel connections, For load effectiveness, α l =0.6, β l =0.4 is the weighting coefficient;

[0123] (4) Overall confidence level C i :

[0124] Calculate the line contribution: E i For node v i The set of connected lines, |E i | Indicates the number of associated lines;

[0125] Determine the dynamic weighting coefficients: The basic weight λ1 = 0.7, and the attenuation coefficient λ2 = 0.1;

[0126] Final overall confidence level: Where α s =0.6 is the structural confidence weight, α f =0.4 is the feature confidence weight.

[0127] Further, in step 5, the target importance of the nodes is calculated to support the selection of important nodes in the power system under conditions of incomplete information. The specific method is as follows:

[0128] Step 5.1: Select the target importance index, including the load impact index ΔP′. L Node degree index K′ i Betweenness centrality index B′ i The target's repair coefficient R′ i Four categories of benefit indicators, and the structural information missingness indicator S miss and feature information missingness index F miss Two types of cost indicators are ranked by their importance using the TOPSIS method to support the selection of power system objectives under conditions of incomplete information.

[0129] (1) Load impact index

[0130] The load impact index assesses the importance of a target by evaluating the amount of load shedding directly or indirectly caused by damage to each target.

[0131]

[0132] Where M is the total equivalent load, N is the total number of power plants, and ΔP L ΔP represents the total load removed from the system after the target is damaged. 1,j Let P be the power flowing from the 1st node to the jth node. G,i Let P represent the power output of the i-th power plant. d,j Let ΔP represent the total load of the j-th equivalent node. d,j This represents the amount of load that needs to be removed from the j-th node when a certain target is damaged;

[0133] If load data is missing, adjust the load impact indicators. Complete the confidence scores for node features;

[0134]

[0135] (2) Node degree index

[0136] The node degree index assesses the importance of each target in the power system by analyzing the degree of each node, and uses an adjacency matrix to describe the relationships between nodes in the power system.

[0137]

[0138] If the structural adjacency matrix of the completed power system network (i.e., the completed power system graph structure) is Then we know the degree K of the i-th node in the network. i satisfy:

[0139]

[0140] The relationship between the degree of a node in a power system network and the number of network edges X is as follows:

[0141]

[0142] If node v i The structure has high missing values ​​and low confidence, meaning its actual number of connections may be underestimated. Therefore, adjustments should be made to the node degree index. To complete the confidence level of the structure;

[0143]

[0144] (3) Betweenness centrality of nodes

[0145] The betweenness centrality of a node is the proportion of the number of paths passing through that node in all shortest paths in a power system network (i.e., the power system graph structure), reflecting the importance of each power target in terms of power transmission.

[0146] node v i Betweenness centrality index B i Defined as:

[0147]

[0148] Where n jk Represents node v j v h The number of paths between them; n jh(i) Represents node v j v h The path passes through node v i The number of;

[0149] If features or structures are missing, the betweenness centrality index is corrected, C i This is the overall confidence level calculated in step 4.

[0150] B′ i =B i ×C i

[0151] (4) Target repair coefficient index

[0152] The target repair coefficient index is evaluated from four perspectives: the region where the power target is located, the type of power target, the rated capacity of the power target, and the difficulty coefficient of the target's maintenance, using a subjective weighting method.

[0153] Target repair coefficient R i The quantitative calculation formula is as follows:

[0154]

[0155] Where a1, a2, and a3 represent the weights of different indicators, the subscript i is used to distinguish different targets, and the subscript j represents different types of power targets: j=1 for power plants, j=2 for substations, and j=3 for power lines; o i This represents the material supply rate for repair, where K rk To analyze the number of dispatch range divisions within the power supply area of ​​this power grid, n k Indicates the k-th element to which the power target belongs. rk Total population of the region; p i U represents the cost of repair. i U represents the rated capacity of the i-th target. j,max L represents the maximum capacity of the third type of target;i L represents the distance of the i-th target from the scheduling center. max q represents the target farthest from the dispatch center within the dispatch area. j (j=1,2,3) represents the target repair coefficients for different types;

[0156] If the node features have a high degree of missing information and low confidence, the uncertainty of the repair time increases, and the repair coefficient needs to be corrected.

[0157]

[0158] Where γ R C is the penalty coefficient, reflecting the sensitivity of missing node information to repair time; i For the overall confidence level calculated in step 4, I miss Overall information missingness α I β I These are the weighting coefficients. The degree of missing structural information. The degree of missing feature information;

[0159] (5) Structural information missingness index:

[0160] Structural information missing degree S miss Defined as the proportion of missing topological connections among nodes, reflecting incomplete structural information;

[0161]

[0162] (6) Feature information missingness index

[0163] The feature information missingness index is the proportion of missing node feature fields, reflecting the incompleteness of attribute information;

[0164] Feature information missing degree F miss definition:

[0165] or

[0166] Among them, X i To complete the features of the preceding node, To complete the node features.

[0167] Step 5.2: Use the TOPSIS method to rank the importance of power targets and obtain the final target importance ranking results for power system nodes.

[0168] Step 5.2.1: Construct the original decision matrix D = [y ij ] n×6n represents the target number of electricity consumption, with each row corresponding to 6 indicators, and m = 6 (including the load impact indicator ΔP′). L Node degree index K′ i Betweenness centrality index B′ i and the target repair coefficient R′ i Four categories of benefit indicators, and the structural information missingness indicator S miss and feature information missingness index F miss Two types of cost indicators).

[0169]

[0170] Normalize matrix D to obtain normalized matrix Z. The normalization formula is as follows:

[0171]

[0172] Step 5.2.2: Assign weights w to each attribute using the entropy weight method. j The weighted normalized decision matrix is ​​constructed as shown in the following equation.

[0173]

[0174] Where x ij =w j ×z ij w represents the weighted, normalized attribute value. j The weight of attribute j.

[0175] Step 5.2.3: Calculate the ideal solution x according to the following formula. + and negative ideal solution x - .

[0176]

[0177] Step 5.2.4: Calculate the power system node v i To the ideal solution x + and negative ideal solution x - The distance is generally measured using the Euclidean norm, as shown in the following formulas.

[0178]

[0179] Step 5.2.5: Calculate the power system node v i Relative proximity R to the ideal solution i As shown in the following formula. R i The larger the value, the greater the importance of the goal, and vice versa, according to R. i The magnitude of the value can be used to rank the importance of nodes in the power system.

[0180]

[0181] A power system target selection system under incomplete information conditions is provided. The system implements the power system target selection method under incomplete information conditions to achieve power system target selection under incomplete information conditions. The system consists of five modules that execute steps 1 to 5 respectively.

[0182] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the power system target selection method under incomplete information conditions, thereby achieving power system target selection under incomplete information conditions.

[0183] A computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, the method for selecting power system targets under conditions of incomplete information is implemented to achieve power system target selection under conditions of incomplete information.

[0184] Compared with existing technologies, the significant advantages of this invention are: 1) It achieves dynamic repair and self-verification through a multi-level comparison graph automatic encoding and decoding network (node ​​feature decoder + structure completion module + subgraph comparison learning), thereby improving the completion accuracy under high missing rate; 2) It designs a power target evaluation index that integrates information missing degree and completion confidence, with information missing degree quantitatively reflecting the degree of original information incompleteness; completion confidence evaluation identifies the credibility of the completion structure and feature information results; and it proposes a dynamic weight function to improve the reliability of the ranking results. Attached Figure Description

[0185] Figure 1 This is a flowchart of the target selection method for power systems under conditions of incomplete information, as described in this invention.

[0186] Figure 2 This is a flowchart of the power target feature structure information completion method based on graph neural networks according to the present invention;

[0187] Figure 3 This is a flowchart of the graph completion learning algorithm based on a multi-level comparison graph automatic encoding and decoding network of the present invention;

[0188] Figure 4 This is a flowchart of the power system target importance ranking method that integrates dynamic compensation according to the present invention;

[0189] Figure 5 This is the topology of the power system in this embodiment. Detailed Implementation

[0190] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0191] A target selection method for power systems under incomplete information conditions is proposed. First, a graph completion learning algorithm is used to dynamically repair fragmented multi-source data. Second, a power system importance assessment model integrating information missingness and completion confidence is proposed to generate a highly reliable target list. The specific steps are as follows:

[0192] Step 1: Obtain relevant data for the power system target, and create a training set containing data on 100 power plants, 30 substations, and 200 transmission lines (including missing data).

[0193] Power plant node data: type code (0-'thermal power', 1-'hydropower', 2-'wind power', 3-'solar power', 4-'nuclear power'), maximum generating capacity (MW), geographical coordinates (latitude and longitude), voltage level (kV).

[0194] Substation node data: type code (5), capacity (MVA), geographical coordinates, voltage level (kV), transformer model code (1-500kV / 2-220kV), turns ratio coefficient (0.9-1.1), number of incoming and outgoing circuits.

[0195] Transmission line data: connection relationship matrix (adjacency matrix A), voltage level (kV), conductor type, conductor starting station, number of parallel circuits, and total regional load (kW).

[0196] Step 2: Using the power system target data collected in Step 1, establish a topological model of the power target system network structure. The topology is described by the set of edges and nodes, modeling the power system as a graph structure G. train =(V train E train ), where V train ={v1,v2,…,v N} represents the set of nodes, including power plants and substations; the set of edges E train ={e ij} represents a power transmission line.

[0197] Node feature matrix X train ∈R N×d N represents the number of nodes, N = 130, and d represents the feature dimension, d = 7.

[0198] Line feature matrix F train ∈R M×6 Each line includes: F train= [Voltage level, conductor type, conductor start point, conductor end point, number of parallel circuits, area load], M is the number of transmission lines, M = 200.

[0199] Adjacency matrix A of power target system train ∈{0,1} N×N If node v i With v j If a physical connection exists (such as a power transmission line), then A ij =1, otherwise 0.

[0200] Step 3: Based on the initial graph structure in Step 2, a graph completion learning algorithm based on a multi-level comparison graph automatic encoding and decoding network is used to supplement the line feature completion part, forming a graph completion learning model for power systems. This model completes the node, line feature, and structural information required for target evaluation, as shown in the attached figure. Figure 3 As shown, the completion confidence of each node and edge is output synchronously.

[0201] Furthermore, step 3 includes the following sub-steps:

[0202] Step 3.1: Input the adjacency matrix A of the power target system train and node feature matrix X train The data are fed into the structure channel and the feature channel, respectively. In the structure channel, the missing graph structure is reconstructed through graph structure diffusion. The reconstructed graph structure is represented as follows: In the feature channels, a trainable parameter matrix is ​​used to initially complete the missing node features. The reconstructed node feature matrix is ​​represented as follows:

[0203] Furthermore, step 3.1 includes the following sub-steps,

[0204] Step 3.1.1: Input the structural adjacency matrix A of the power target system train A personalized PageRank algorithm is used to analyze the structural adjacency matrix A. train Structural diffusion is performed, with physical constraints imposed on the power system structure during diffusion. These constraints include prohibiting direct connections between power plant nodes and implementing voltage level matching constraints, requiring connections only between nodes of the same voltage level. The parameters are set as follows: restart probability α = 0.15, iteration count T = 10. Sparsity is achieved by retaining the three strongest connections in the Top-K network, generating a complete adjacency matrix.

[0205] The diffusion formula is shown below:

[0206]

[0207] in Let A be the adjacency matrix at time t. trainα is the restart probability, I n It is an identity matrix with the same dimensions as the adjacency matrix, initially

[0208] A is removed using k-nearest neighbor sparsification technique SR For unimportant edges, generate a complete adjacency matrix. Its elements Represented as:

[0209]

[0210] Where kNN(i) represents the set of the top (k) neighbors of the i-th node, if node v j Belongs to node v i The top-k neighbor set (i.e., j∈kNN(i)) is then retained. Otherwise, set to zero. T = 10 is the number of iterations, A SR The remaining elements are 0, and the diffusion result of each node retains the 3 strongest connections of the node.

[0211] Step 3.1.2: Input the node feature matrix X train ∈R N×d Perform node feature matrix initialization and completion, and construct a matrix with X. train Parameter matrices Θ of identical size ij ∈R N×d N is the number of nodes, d = 7, used to fill in X. train The missing values ​​in the matrix. Let the feature matrix of the completed nodes be... It can be represented as:

[0212]

[0213] Where X i,j For node v i The j-th dimension feature, when X i,j Use Θ when missing i,j The values ​​are used to fill the gaps, and physical constraints are applied to the node characteristics during the filling process (parameters related to the shielded transformer of power plant nodes, and fixed type parameters of substation nodes). Θ ij The parameters are randomly initialized and bound to the neural network model, and then iteratively optimized using an optimizer.

[0214] Step 3.2: Complete the structural adjacency matrix in Step 3.1.1. The data is fed into a structural encoder, and positional encoding is used as the initial feature to compute the node embedding representation Z of the structural channel. SR The completed structural adjacency matrix The adjacency matrix for guided feature diffusion is obtained by weighted summation. The node features are fed into the feature channel to guide feature diffusion. The diffused node features are then fed into the feature encoder to compute the node embedding representation Z of the feature channel. FR .

[0215] Furthermore, step 3.2 includes the following sub-steps,

[0216] Step 3.2.1: Construct a structure encoder based on a graph convolutional network, taking the structure adjacency matrix completed in Step 3.1.1 as input. The positional encoding (PE) is used, and the structure encoding is performed based on a two-layer graph convolutional neural network (GCN). The output is Z. SR .

[0217] For node data V in step 2 train The location information in the data is modeled, and a location code PE∈R is generated for each node. N×d ,

[0218]

[0219] Where PE(v) i ) for a single node v i Position encoding vector, W∈R N×d It is a learnable weight matrix, b∈R N×d For learnable bias vectors, It is node v i The one-hot encoding is an N-dimensional vector.

[0220] By using a two-layer graph convolutional neural network (GCN), the encoder output of the structural channel is obtained as the node embedding representation Z of the structural channel. SR .

[0221]

[0222] in This represents the position encoding vector for each node, where σ is a non-linear activation function. This represents the weight matrix of the first and second graph convolutional layers.

[0223] Step 3.2.2: Construct a feature channel encoder based on feature diffusion and multilayer perceptron, and complete the structural adjacency matrix in Step 3.1.1. The adjacency matrix for guided feature diffusion is obtained by weighted summation. The data is fed into the feature channel to guide feature diffusion. The diffused node features are then... The data is fed into a feature encoder, and a multilayer perceptron (MLP) is used to compute the node embedding representation Z of the feature channels. FR .

[0224]

[0225] Where Λ∈R N×N It is a learnable coefficient matrix. To guide the adjacency matrix of feature diffusion, by It is obtained by weighted multiplication with the coefficient matrix.

[0226] Input the node feature matrix after completion in step 3.1.2. Perform node feature diffusion.

[0227]

[0228] Where X (0) The node feature matrix representing the initial time step, which is the encoder input node feature matrix. X (t) Let represent the node feature matrix at time t, α∈(0,1] represent the restart probability, here α=0.3, T represent the number of rounds of diffusion propagation, T=5, t∈[0,T-1], This represents the node feature representation of the output after T rounds of diffusion propagation.

[0229] In the message transformation step, a two-layer MLP module is introduced to represent the feature representations of the aggregated node's neighbors. The node feature matrix after diffusion propagation is then used. The data is fed into an MLP encoder to obtain the feature channel node embedding representation Z. FR The message conversion process is as follows:

[0230]

[0231] Where W1 and W2 are the learnable weight matrices of the first and second layers of the MLP, respectively, b1 and b2 are the corresponding bias matrices, σ is the nonlinear activation function (the same as the nonlinear activation function in step 3.2.1), and Z... FR H is the node embedding representation of the learned feature channels, and H is the intermediate representation of the node features in the feature channels after the first layer transformation.

[0232] Step 3.3: Input the node embedding representation Z of the structural channel from Step 3.2.1 SR The node embedding representation Z of the feature channels in step 3.2.2 FR Using attention mechanism to fuse Z FR and Z SR The weighted fused node embedding representation Z is then processed by a softmax classifier to obtain the classification result.

[0233] Furthermore, step 3.3 includes the following sub-steps,

[0234] Step 3.3.1: Calculate the node embedding representation Z of the feature channelsFR Feature channel weights γ f The node embedding representation Z of the computational structure channel. SR Structure channel weight γ s .

[0235]

[0236] in and Represents the linear transformation MLP, and W represents the parameters of the MLP. F and W′ F Let γ be the learnable weight matrix, tanh be the nonlinear activation function for calculating the weights of the feature and structural channels, and γ be the weight matrix. s and γ f These represent the raw scores of the feature channel and the structure channel, respectively, calculated initially through the attention mechanism.

[0237] Step 3.3.2: Then, the weighting coefficients are normalized according to γ. s and γ f Calculate the weight coefficients of the two channels after normalization.

[0238]

[0239] Where w s and w f These represent the weights of the feature channel and the structure channel, respectively.

[0240] Step 3.3.3: Finally, calculate the weighted fused node embedding representation Z.

[0241] Z = w f ×Z FR +w s ×Z SR

[0242] Step 3.3.4: After obtaining the weighted fusion node embedding representation Z in step 3.3.3, use a fully connected layer to reduce the dimension of Z to the category dimension, thereby obtaining the score S of the sample node belonging to each category.

[0243] S = fc(Z)

[0244] Where S∈R n×c c represents the number of categories, s i,j Let represent the score of the i-th node belonging to the j-th class, and fc denotes a fully connected layer. Using...

[0245] Finally, the category prediction matrix of the nodes is obtained. in This represents the probability that the i-th node belongs to the j-th class. The cross-entropy loss function is then used to calculate... The error value between the actual label Y and the true label Y is used to train the graph completion learning model for power systems in this patent.

[0246] Step 3.4: Embed the node representation Z of the feature channels FR The data is fed into the feature decoder to reconstruct the node features. Embedding the nodes of the structural channel to represent Z SR The data is fed into the feature decoder to reconstruct the node structure. Used for subsequent calculation of contrast loss for alignment. and

[0247] Furthermore, step 3.4 includes the following sub-steps,

[0248] Step 3.4.1: Input the node embedding representation Z of the structural channel from Step 3.2.1 SR The node embedding representation Z of the feature channels in step 3.2.2 FR , for Z FR and Z SR Apply a random mask.

[0249]

[0250] Among them, MASK z Z represents a mask matrix, where the elements mask∈{0,1}. FR and Z SR These represent the node embedding representations of the feature channel and the structure channel, respectively. Z is the node embedding representation of the feature channels. FR The result after masking. It is a structural channel node embedding representation Z SR The result after masking, MASK z With Z FR and Z SR They have the same dimensions.

[0251] Step 3.4.2: Use a single MLP layer to implement decoding, and convert Z... FR Input the feature decoder, reconstruct the feature channel node representation, and compute the reconstructed features of the feature channels. Z SR Input the feature decoder, reconstruct the structural channel node representation, and compute the reconstructed features of the structural channels.

[0252]

[0253] Where ψ fMLP decoder for feature channels Parameter, ψ s MLP decoder for structured channels parameter.

[0254] Step 3.5: After completing the structure, perform a process analysis on the line feature F. train ∈R M×6 To complete the information, the line characteristics include six features: voltage level, conductor type, conductor start and end points, number of parallel circuits, and regional load.

[0255] Furthermore, step 3.5 includes the following sub-steps,

[0256] Step 3.5.1: Perform voltage level completion for each completed line e ij According to the voltage levels V at both ends i and V j Determine the line voltage level:

[0257] V e =max(V i V j )

[0258] Where V e For line e ij The voltage level is determined by taking the higher voltage value at both ends as the line voltage level. If the node voltage is unknown, it is set to a preset value to ensure that the voltage requirement is met. e >35KV.

[0259] Step 3.5.2: Complete the wire type selection based on voltage level V. e Match the preset model library.

[0260]

[0261] Where X e For the wire type, P(m|V) e ) indicates a known voltage level V e The probability of occurrence of type m, V e =500KV model is LGJ-400, V e =300KV model LGJ-300, V e =250KV model LGJ-240, V e =110KV model LGJ-150.

[0262] Step 3.5.3: Complete the start and end points of the traverse lines, and then use the completed adjacency matrix. Extract the physical connection relationships and determine the line endpoints.

[0263] like Then a line e is generated. ij If node v i For power plants (node ​​v) i If the type encoding is 0, 1, 2, 3, 4, then v i As the starting point of the route S e v j The endpoint is v; i , and v j All are substations (node ​​v) i The type code is 5). Compare the sum of the longitude and latitude coordinates of two nodes, and select the node with the smaller sum as the starting point. Its geographical sorting logic is East → West (the larger the longitude, the further east it is, and the smaller the longitude, the further west it is) and North → South (the larger the latitude, the further north it is, and the smaller the latitude, the further south it is) to determine the starting point.

[0264] Step 3.5.4: Complete the parallel circuit count based on the total regional load L. region Predict the number of parallel circuits P e .

[0265]

[0266] Where β P The regression coefficients are obtained by fitting historical data to ensure P0. e ≥1. L region The total load of the region is defined by clip(x,a,b), which limits x to the range [a,b].

[0267] Step 3.5.5: Perform regional load supplementation, and supplement the current regional load by weighted average of the loads of adjacent regions.

[0268]

[0269] Where L e Let N(e) be the area load of line e, and N(e) be the set of lines adjacent to line e, with weights... dist(e,k) is the geographical distance between lines e and k, ∈ is a minimal constant, and L k The known load value of the adjacent line k.

[0270] Step 3.6: Calculate the classification results Cross-entropy loss with the real data label Y The embedding representation of node i in the feature channel Z FR The node i embedded in the structural channel represents Z. SR encoder contrast loss Reconstruction features of feature channels and the reconstruction features of structural channels Decoder contrast loss And the total loss is obtained by weighted summation.

[0271] Furthermore, step 3.6 includes the following sub-steps,

[0272] Step 3.6.1: Calculate the classification results Cross-entropy loss with the real data label Y

[0273]

[0274] Where Y i This represents the true label of the i-th sample. Let represent the predicted label of the i-th sample, and n represent the number of labels.

[0275] Step 3.6.2: Calculate type alignment loss during the encoding phase. Supervision signals are extracted from unlabeled samples to avoid overfitting.

[0276]

[0277] in Let τ be the prototype vector of category j in the feature channel and the structure channel, τ be an adjustable scaling factor, and C be the total number of categories.

[0278]

[0279] Where μ j Let be the prototype vector of category j, where the superscript indicates the feature channel FR or the structural channel SR. The predicted categories for node i are represented by c, where c is the total number of categories, and q is the number of categories. i Represents node v i The weights, Q i The weight x for all nodes predicted as class j i the sum of Filter all nodes v with predicted category j i , For node v of the structural channel in step 3.2.1 i Embedded representation Z SR , For the node v of the feature channel in step 3.2.2 i Embedded representation Z FR ;

[0280] Step 3.6.3: Compare the loss using the MLP decoder for the feature channel and the MLP decoder for the structure channel as described in step 3.4.2. Supervise the reconstruction process and use contrastive loss alignment. and Calculate the reconstruction contrast loss

[0281]

[0282] in For decoder reconstruction loss, The reconstructed features of the feature channels and structural channels in step 3.5.2 are denoted as , where is the vector inner product, N is the total number of nodes, and τ is the adjustment coefficient.

[0283] Step 3.6.4: Calculate the total loss function using weighted summation.

[0284]

[0285] Where λ1 = 0.7 is the encoder loss. The weighting coefficients, λ² = 0.3, represent the decoder loss. The weighting coefficients.

[0286] Step 3.6: Update the parameters of the graph completion learning model for the power system using the gradient backpropagation algorithm until convergence.

[0287] Step 4: Conduct experiments based on test data, and use the graph completion learning model for power systems trained in Step 3 to complete the structure and node and line features.

[0288] Step 4.1: The test data consists of a power system comprising 2 power plants, 3 substations, and 5 transmission lines. The characteristic parameters of each node are shown in the table below. Nan represents missing data.

[0289] The power plant data is as follows:

[0290] Serial Number Type Encoding Maximum generating capacity (MW) Geographic coordinates Voltage level (kV) 3 0 500 120.1302,33.3795 500 4 1 Nan 116.7432,37.7919 500

[0291] The input substation data is as follows:

[0292]

[0293] The input transmission line data is as follows:

[0294] starting point end Voltage level (kV) wire type Number of parallel circuits Total regional load (kW) 0 2 500 LGJ-400 2 200 1 3 500 LGJ-400 1 150 2 3 220 LGJ-240 1 100 3 4 Nan LGJ-150 1 50

[0295] Step 4.2: Model the test set power system as a graph structure G test =(V test E test ), constructing a topology model such as Figure 5 As shown, construct the feature matrix X. test ∈R N×dWhere N=5 is the total number of nodes, d=7 is the feature dimension, and missing values ​​are marked as NaN. Construct the structural adjacency matrix A. test ∈{0,1} N×N There exists a physical connection with a value of 1. Construct the line feature matrix F∈R. M×6 M = 4.

[0296] Step 4.3: Perform structural completion, input A test Through diffusion and constraint in step 3, the following is generated: The physical constraints of the power system structure include prohibiting direct connections between power plant nodes and applying voltage level matching constraints, requiring that only nodes of the same voltage level are allowed to connect. The output structure completion result is: edge (2,4).

[0297] Step 4.4: Perform feature completion, input X test Generate through step 3 The physical constraints for node characteristics are the shielded transformer-related parameters for power plant nodes (ID0-1) (columns 5-7) and the fixed type parameters for substation nodes (ID 2-4) (column 0). The output node characteristics are:

[0298]

[0299] Step 4.5: Based on the completed structure and node features, calculate the line features of edges (2,4) and (3,4) according to the calculation formula in Step 3.5, as shown in the table below.

[0300] starting point end Voltage level (kV) wire type Number of parallel circuits Total regional load (kW) 2 4 220 LGJ-240 1 80 3 4 110 LGJ-150 1 50

[0301] Step 4.5: After graph completion, calculate the confidence scores for structural completion, node feature completion, line feature completion, and overall confidence score.

[0302] (1) Confidence of structural completion

[0303] The confidence level of structural completion reflects the reliability of topology completion, and the computation node v i The average connection probability with all other nodes; a higher value indicates more reliable topology completion.

[0304]

[0305] Where A comp ∈R N×N This is the completed probabilistic adjacency matrix. Represents node v i and v j The connection probability, where N is the total number of nodes;

[0306] The node embedding representation Z of the structural channel output by the structural encoder (GCN) in step 3.2.1 SR Generate the completed probabilistic adjacency matrix using MLP:

[0307]

[0308] Where θ is the Sigmoid function, and its output value is... This represents the probability of node connection.

[0309] (2) Confidence of node feature completion

[0310] The confidence score of node feature completion is used to evaluate the quality of node feature repair. Based on the absolute error between the known node features and the completed node features, it is normalized and then mapped by tanh. Finally, it is converted into the confidence score of node feature completion by 1-tanh(·).

[0311]

[0312] in X is the completed node feature matrix. known Given the known node feature values, M∈{0,1} N×D The missing mask matrix is ​​denoted by ∈, where ∈ is the smoothing coefficient (usually taken as 1e). -6 ), where d is the number of feature dimensions, d = 7.

[0313] (3) Confidence of line completion features

[0314] The confidence level of line feature completion is calculated by weighting the connection probability (structural reliability) of the two end nodes and the verification results of physical constraints (functional rationality).

[0315]

[0316] in Let S be the node connection probability of nodes s and t at both ends of line e in the completed adjacency matrix. For voltage effectiveness, For the validity of the number of parallel connections, For load effectiveness, α l =0.6, β l =0.4 is the weighting coefficient.

[0317] (4) Overall confidence level C i :

[0318] A dynamic weighting mechanism is constructed to calculate the overall confidence score and dynamically balance the confidence scores of the node's own structure. Feature confidence Contribution to associated lines

[0319] Line contribution: E i For node v i The set of connected lines, |E i | Indicates the number of associated lines;

[0320] Dynamic weighting coefficients: The basic weight λ1 = 0.7, and the attenuation coefficient λ2 = 0.1;

[0321] Final overall confidence level:

[0322] Where α s =0.6 is the structural confidence weight, α f =0.4 represents the feature confidence weight.

[0323] The results of each confidence level calculation are shown in the table below:

[0324] node Confidence of structural completion Node feature completion confidence Overall confidence level 0 1.000 0.856 0.965 1 1.000 0.768 0.944 2 0.830 0.882 0.855 3 1.000 1.000 1.000 4 0.700 0.827 0.766

[0325]

[0326]

[0327] Step 5: Based on the completed power system node features, line features, and structural information from Step 4, the target importance ranking method is used to sort the 2 power plants and 3 substations in the test data.

[0328] Furthermore, step 5 includes the following sub-steps:

[0329] Step 5.1: Based on the power target system diagram structure from Step 4, the confidence level from Step 4.5, and the completed node features, line features, and structural information, calculate the target importance index based on the topology model from Step 4.2, mainly including the load impact index ΔP′. L Node degree index K′ i Betweenness centrality index B′ and target repair coefficient R′ i Four categories of benefit indicators, and the structural information missingness indicator S miss and feature information missingness index F miss Two types of cost indicators.

[0330] (1) Load impact index

[0331] The importance of a target is assessed by evaluating the amount of load shear caused directly or indirectly by damage to each target.

[0332]

[0333] Where M is the total equivalent load, N is the total number of power plants, and ΔP L ΔP represents the total load removed from the system after the target is damaged. 1,j Let P be the power flowing from the 1st node to the jth node. G.i Let P represent the power output of the i-th power plant. d,j Let ΔP represent the total load of the j-th equivalent node. d,j This represents the amount of load that needs to be removed from the j-th node when a certain target is damaged.

[0334] If load data is missing, adjust the load impact indicators. Complete the confidence level for the feature.

[0335]

[0336] (2) Node degree index

[0337] The importance of each target in the power system can be quickly assessed by analyzing the degree of each node in the network, and the relationship between the nodes in the power system is described by an adjacency matrix.

[0338]

[0339] Figure 5 The correlation matrix of the topology graph shown is

[0340]

[0341] If the structural adjacency matrix of the completed power system network (i.e., the completed power system graph structure) Then we know the degree K of the i-th node in the network. i satisfy:

[0342]

[0343] The relationship between the degree of a node in a power system network and the number of network edges X is as follows:

[0344]

[0345] If node v i The structure has high missing values ​​and low confidence, meaning its actual number of connections may be underestimated. Therefore, adjustments should be made to the node degree index. To complete the confidence level of the structure;

[0346]

[0347] (3) Betweenness centrality of nodes

[0348] The betweenness centrality of a node is the proportion of the number of paths passing through that node in all shortest paths in a power system network (i.e., the power system graph structure), reflecting the importance of each power target in terms of power transmission.

[0349] node v i Betweenness centrality index B i Defined as:

[0350]

[0351] Where n jk Represents node v j v h The number of paths between them; n jh(i) Represents node v j v h The path passes through node v i The number of;

[0352] If features or structures are missing, the betweenness centrality index is corrected, C i This is the overall confidence level calculated in step 4.

[0353] B′ i =B i ×C i

[0354] (4) Target repair coefficient index

[0355] The assessment of power targets is conducted using a subjective weighting method, considering factors such as the target's location, type, rated capacity, and maintenance difficulty. The target repair coefficient R is... i The quantitative calculation formula is as follows.

[0356]

[0357] Where a1, a2, and a3 represent the weights of different indicators, the subscript i is used to distinguish different targets, and the subscript j represents different types of power targets: j=1 for power plants, j=2 for substations, and j=3 for power lines; o i This represents the material supply rate for repair, where K rk To analyze the number of dispatch range divisions within the power supply area of ​​this power grid, n k Indicates the k-th element to which the power target belongs. rk Total population of the region; p i U represents the cost of repair. i U represents the rated capacity of the i-th target. j,max L represents the maximum capacity of the third type of target; i L represents the distance of the i-th target from the scheduling center.max q represents the target farthest from the dispatch center within the dispatch area. j Indicates the target repair coefficient for different types;

[0358] If the node features have a high degree of missing information and low confidence, the uncertainty of the repair time increases, and the repair coefficient needs to be corrected.

[0359]

[0360] Where γ R C is the penalty coefficient, reflecting the sensitivity of missing node information to repair time; i For the overall confidence level calculated in step 4, I miss Overall information missingness α I β I These are the weighting coefficients. The degree of missing structural information. The degree of missing feature information;

[0361] (5) Structural information missingness index:

[0362] Structural information missing degree S miss Defined as the proportion of missing topological connections among nodes, it reflects the incompleteness of structural information (cost-based indicator).

[0363]

[0364] (6) Feature information missingness index

[0365] The feature information missingness index is the proportion of missing feature fields in a node, reflecting the incompleteness of attribute information (a cost-based indicator). Feature Information Missingness F miss definition:

[0366] or

[0367] Among them, X i To complete the features of the preceding node, To complete the node features.

[0368] Step 5.2: Use the TOPSIS method to rank the importance of power targets and obtain the final target importance ranking results for power system nodes.

[0369] Step 5.2.1: Construct the original decision matrix D = [y ij ] n×6 n represents the number of electricity targets (n=5), each row corresponds to 6 indicators, and m=6 (including the load impact indicator ΔP′). L Node degree index K′i Betweenness centrality index B′ i and the target repair coefficient R′ i Four categories of benefit indicators, and the structural information missingness indicator S miss and feature information missingness index F miss Two types of cost indicators).

[0370]

[0371] Normalize matrix D to obtain normalized matrix Z. The normalization formula is as follows:

[0372]

[0373] Step 5.2.2: Assign weights w to each attribute using the entropy weight method. j The weighted normalized decision matrix is ​​constructed as shown in the following equation.

[0374]

[0375] Where x ij =w j ×z ij w represents the weighted, normalized attribute value. j w is the weight of attribute j. j = [0.25, 0.2, 0.25, 0.1, 0.1, 0.1].

[0376] Step 5.2.3: Calculate the ideal solution x according to the following formula. + and negative ideal solution x - .

[0377]

[0378] Step 5.2.4: Calculate the power system node v i To the ideal solution x + and negative ideal solution x - The distance is generally measured using the Euclidean norm, as shown in the following formulas.

[0379]

[0380] Step 5.2.5: Calculate the power system node v i Relative proximity R to the ideal solution i As shown in the following formula. R i The larger the value, the greater the importance of the goal, and vice versa.

[0381]

[0382] According to R i The magnitude of the value can be used to rank the importance of nodes in the current power system as follows:

[0383] node 0 1 2 3 4 TOPSIS sorting results 4 3 1 2 5

[0384] In summary, this invention completes incomplete information in power systems by using graph completion learning technology based on an improved multi-level comparison graph automatic encoding and decoding network. This makes existing evaluation methods based on complete information widely applicable and more reliable than traditional evaluation methods that use graph neural networks to complete information.

[0385] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0386] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these modifications and improvements all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for selecting targets in a power system under conditions of incomplete information, characterized in that, Includes the following steps: Step 1: Obtain target-related data of the power system as a training set, which includes power plant node data, substation node data and transmission line data; Step 2: Construct a set of nodes using power plants and substations, and a set of edges using transmission line data. Based on this, establish a topological model of the power target system network structure, model the power system as a graph structure, and then determine the node feature matrix, line feature matrix, and adjacency matrix. Step 3: Based on the graph completion learning algorithm of the multi-level contrast graph automatic encoding and decoding network, in the structural channel, the missing graph structure is reconstructed by graph structure diffusion. In the feature channel, the missing node features are initially completed by a trainable parameter matrix. At the same time, the node embedding representations of the structural channel and the feature channel are calculated, weighted and fused, and then passed through a classifier to obtain the classification result. The line features are then completed to form a graph completion learning model for power systems. Step 4: For the target power system, perform structural completion, node feature completion, and line feature completion based on the trained graph completion learning model for power systems, and calculate the confidence scores of structural completion, node feature completion, line feature completion, and overall confidence score. Step 5: Calculate the target importance of power system nodes to support power system target selection under conditions of incomplete information.

2. The power system target selection method under incomplete information conditions according to claim 1, characterized in that, Step 1: Obtain target-related data of the power system as the training set, which includes power plant node data, substation node data, and transmission line data, wherein: The power plant node data includes type code, maximum generating capacity, geographic coordinates, and voltage level. The type code values ​​and corresponding types are 0-'thermal power', 1-'hydro power', 2-'wind power', 3-'solar power', and 4-'nuclear power'. Substation node data includes type code, capacity, geographic coordinates, voltage level, transformer model code, turns ratio coefficient, and number of incoming and outgoing line circuits. The type code value is 5. Transmission line data includes a connection matrix, as well as voltage level, conductor type, conductor starting station, number of parallel circuits, and total regional load.

3. The power system target selection method under incomplete information conditions according to claim 2, characterized in that, Step 2: Construct a node set using power plants and substations, and an edge set using transmission line data. Based on this, establish a topological model of the power target system network structure, modeling the power system as a graph structure. Then, determine the node feature matrix, line feature matrix, and adjacency matrix. The specific method is as follows: Step 2.1: Construct a set of nodes using power plants and substations, and a set of edges using transmission line data. Based on this, establish a topological model of the power target system network structure, modeling the power system as a graph structure, represented as: G train =(V train ,E train ) Where V train ={v1,v2,…,v N } represents the set of nodes, including power plants and substations; the set of edges E train ={e ij } represents a power transmission line; Step 2.2, determine the node feature matrix, the line feature matrix, and the adjacency matrix, where: Node feature matrix X train ∈R N×d N represents the number of nodes, and d represents the feature dimension; Line feature matrix F train ∈R M×6 Each line includes: F train = [Voltage level, conductor type, conductor start point, conductor end point, number of parallel circuits, area load], where M is the number of transmission lines; Adjacency matrix A train ∈{0,1} N×N If node v i With v j If a power transmission line exists, then A ij =1, otherwise 0.

4. The power system target selection method under incomplete information conditions according to claim 1, characterized in that, Step 3: Based on a multi-level contrast graph automatic encoding / decoding network, a graph completion learning algorithm is used. In the structural channel, missing graph structures are reconstructed through graph structure diffusion. In the feature channel, a trainable parameter matrix is ​​used to initially complete missing node features. Simultaneously, the node embedding representations of the structural and feature channels are calculated, weighted, fused, and then passed through a classifier to obtain the classification result. The line features are then fully processed to form a graph completion learning model for power systems. The specific method is as follows: Step 3.1: Input the structural adjacency matrix A train and node feature matrix X train The graphs are fed into the structure channel and the feature channel, respectively. In the structure channel, the missing graph structure is reconstructed through graph structure diffusion. The reconstructed graph structure is represented as the completed structure adjacency matrix. In the feature channels, a trainable parameter matrix is ​​used to initially complete the missing node features. The reconstructed node feature matrix is ​​represented as follows: Step 3.1.1: Input the structural adjacency matrix A train A personalized PageRank algorithm is used to analyze the structural adjacency matrix A. train Perform structural diffusion and set physical constraints on the power system structure during the diffusion process, including prohibiting direct connections between power plant nodes and allowing connections only between nodes of the same voltage level, to generate a completed structural adjacency matrix. in: Structural diffusion formula: in Let A be the structural adjacency matrix at time t. train α is the restart probability, I n It is an identity matrix with the same dimensions as the structural adjacency matrix, initially A is removed using k-nearest neighbor sparsification technique SR For unimportant edges, generate the completed structural adjacency matrix. Its elements for: Where kNN(i) represents the set of the top (k) neighbors of the i-th node, if node v j Belongs to node v i The top-k neighbor set is then retained. Otherwise, set to zero; Step 3.1.2: Input the node feature matrix X train ∈R N×d Perform node feature matrix initialization and completion; Building with X train Parameter matrices Θ of identical size ij ∈R N×d N is the number of nodes used to fill in missing values ​​in the node feature matrix. Let the initialized and completed node feature matrix be... This is represented as: Where X i,j For node v i The j-th dimension feature, when X i,j Use random parameter Θ when missing i,j The values ​​are used to fill the values, and physical constraints are applied to the node characteristics during the filling process, including parameters related to the shielded transformer of the power plant node and the fixed type parameters of the substation node. Step 3.2: Complete the structural adjacency matrix. The data is fed into a structural encoder, and positional encoding is used as the initial feature to compute the node embedding representation Z of the structural channel. SR ; Complete the structural adjacency matrix The adjacency matrix for guided feature diffusion is obtained by weighted summation. The node features are fed into the feature channel to guide feature diffusion. The diffused node features are then fed into the feature encoder to compute the node embedding representation Z of the feature channel. FR ; Step 3.2.1: Construct a structure encoder based on a graph convolutional network, taking the completed structure adjacency matrix as input. The positional encoding (PE) is used, and the structure encoding is performed based on a two-layer graph convolutional neural network (GCN). The output is Z. SR ; For the node set V train The location information in the data is modeled, and a location code PE∈R is generated for each node. N×d , Where PE(v) i ) for a single node v i Position encoding vector, W∈R N×d It is a learnable weight matrix, b∈R N×d For learnable bias vectors, It is node v i The one-hot encoding is an N-dimensional vector; By using a two-layer graph convolutional neural network (GCN), the encoder output of the structural channel is obtained as the node embedding representation Z of the structural channel. SR ; in This represents the position encoding vector for each node, where σ is a non-linear activation function. This represents the weight matrix of the first and second graph convolutional layers; Step 3.2.2: Construct a feature channel encoder based on feature diffusion and multilayer perceptron, and use the completed structural adjacency matrix. The adjacency matrix for guided feature diffusion is obtained by weighted summation. The data is fed into the feature channel to guide feature diffusion, and the diffused node features are then... The data is fed into a feature encoder, and a multilayer perceptron (MLP) is used to compute the feature channel node embedding representation Z. FR ; Where Λ∈R N×N It is a learnable coefficient matrix. To guide the adjacency matrix of feature diffusion, by It is obtained by weighted multiplication with the coefficient matrix; Input the completed node feature matrix Perform node feature diffusion; Where X (0) X represents the node feature matrix at the initial time step. (t) The matrix represents the node feature matrix at time t, where T represents the number of rounds of diffusion propagation. This represents the node feature matrix output after T rounds of diffusion propagation, where α∈(0,1] is the restart probability; A two-layer MLP feature encoder module is introduced to perform message transformation, converting the node feature matrix after diffusion propagation. The data is fed into an MLP feature encoder to obtain the feature channel node embedding representation Z. FR The message conversion process is as follows: Where W1 and W2 are the learnable weight matrices of the first and second layers of the MLP, respectively, b1 and b2 are the corresponding bias matrices, σ is the nonlinear activation function, and Z... FR H is the node embedding representation of the learned feature channels, where H is the intermediate representation of the node features in the feature channels after the first layer transformation. Step 3.3: Input the node embedding representation Z of the structural channel from Step 3.2.1 SR The node embedding representation Z of the feature channels in step 3.2.2 FR Using attention mechanism to fuse Z FR and Z SR The weighted fused node embedding representation Z is then processed by a softmax classifier to obtain the classification result. Step 3.3.1: Calculate the node embedding representation Z of the feature channels FR Feature channel weights γ f The node embedding representation Z of the computational structure channel. SR Structure channel weight γ s ; in and Represents the linear transformation MLP, and W represents the parameters of the MLP. F and W′ F Let γ be the learnable weight matrix, tanh be the nonlinear activation function for calculating the weights of the feature and structural channels, and γ be the weight matrix. s and γ f These represent the raw scores of the feature channel and the structure channel, respectively, calculated initially through the attention mechanism; Step 3.3.2: Adjust the feature channel weights γ f and structural channel weight γ s After normalization, the weight coefficients of the two channels are: Where w s and w f These represent the weights of the feature channel and the structure channel after normalization, respectively. Step 3.3.3: Calculate the weighted fused node embedding representation Z; Z=w f ×Z FR +in s ×Z SR Step 3.3.4: After obtaining the weighted fusion node embedding representation Z in step 3.3.3, use a fully connected layer to reduce the dimension of Z to the category dimension, thereby obtaining the score S of the sample node belonging to each category; S = fc(Z) Where S∈R n×c c represents the number of categories, s i,j Let fc represent the score of the i-th node belonging to the j-th class, and fc represent the fully connected layer. The softmax function is used to map the score to the (0,1) interval, and make the sum of the probabilities of the node belonging to each class equal to 1. Finally, the category prediction matrix of the nodes is obtained. in This represents the probability that the i-th node belongs to the j-th class; Step 3.4: Analyze the line characteristic F train ∈R M×6 To complete the information, the line characteristics include six features: voltage level, conductor type, conductor start and end points, number of parallel circuits, and area load. Step 3.4.1: Perform voltage level completion for each completed line e ij According to the voltage levels V at both ends i and V j Determine the line voltage level: V e =max(V i ,V j ) Where V e For line e ij The voltage level is determined by taking the higher voltage value at both ends as the line voltage level. If the node voltage is unknown, it is set to a preset value to ensure that the voltage requirement is met. e >35KV; Step 3.4.2: Complete the wire type based on voltage level V. e Match the preset model library; Where X e For the wire type, P(m|V) e ) indicates a known voltage level V e The probability of occurrence of type m, V e =500KV model is LGJ-400, V e =300KV model LGJ-300, V e =250KV model LGJ-240, V e = 110KV model LGJ-150; Step 3.4.3: Complete the start and end points of the traverse, and then use the completed adjacency matrix. Extract physical connections and determine line endpoints; like Then a line e is generated. ij If node v i If it is a power plant, then v i As the starting point S of the route e v j The endpoint is v; i and v j Both are substations. The sum of the longitude and latitude coordinates of the two nodes is compared, and the node with the smaller sum is selected as the starting point. The geographical sorting logic is East → West and North → South to determine the starting point. Step 3.4.4: Complete the parallel circuit count based on the total regional load L. region Predict the number of parallel circuits P e ; Where β P For the regression coefficients, ensure P e ≥1, L region The total load of the region is defined, and clip(x,a,b) restricts x to the range [a,b]. Step 3.4.5: Perform regional load supplementation, using a weighted average of adjacent regional loads to supplement the current regional load; Where L e Let N(e) be the area load of line e, and N(e) be the set of lines adjacent to line e, with weights... dist(e,k) is the geographical distance between lines e and k, ∈ is a minimal constant, and L k The known load value of the adjacent line k.

5. The power system target selection method under incomplete information conditions according to claim 4, characterized in that, The graph completion learning model for power systems is trained using the cross-entropy loss function. The specific method is as follows: Step 3.5: Embed the node representation Z of the feature channels FR The data is fed into the feature decoder, and the node channels are used to reconstruct the features. Embedding the nodes of the structural channel to represent Z SR The data is fed into the structure decoder to calculate the structural channel reconstruction features. Used for subsequent calculation of contrast loss for alignment. and Step 3.5.1: Input the node embedding representation Z of the structural channel from Step 3.2.1 SR The node embedding representation Z of the feature channels in step 3.2.2 FR , for Z FR and Z SR Perform random masking; Among them, MASK z Let represent a mask matrix, where the elements mask∈{0,1}. Z is the node embedding representation of the feature channels. FR The result after masking. It is a structural channel node embedding representation Z SR The result after masking, MASK z With Z FR and Z SR They have the same dimensions; Step 3.5.2: Use a single MLP layer to implement decoding, converting Z... FR Input to the feature decoder to compute the reconstructed features of the feature channels. Z SR Input to the structure decoder to compute the reconstructed features of the structure channels. Where ψ f MLP decoder for feature channels Parameter, ψ s MLP decoder for structured channels parameter; Step 3.6: Calculate the classification results Cross-entropy loss with the real data label Y Node embedding representation of feature channels Z FR The node embedding representation of the structural channel Z SR encoder contrast loss Reconstruction features of feature channels and the reconstruction features of structural channels Decoder contrast loss And the total loss is obtained by weighted summation. Step 3.6.1: Calculate the classification results Cross-entropy loss with the real data label Y Where Y i This represents the true label of the i-th sample. Let represent the predicted label of the i-th sample, and n represent the number of labels; Step 3.6.2: Calculate the encoder alignment loss during the encoding stage. Extract supervision signals from unlabeled samples to avoid overfitting; in Let τ be the prototype vector of category j in the feature channel and the structure channel, τ be an adjustable scaling factor, and C be the total number of categories; Where μ j Let be the prototype vector of category j, where the superscript indicates the feature channel FR or the structural channel SR. The predicted categories for node i are represented by c, where c is the total number of categories, and q is the number of categories. i Represents node v i The weights, Q i The weights x for all nodes predicted as class j i The sum of Filter all nodes v with predicted category j i , For node v of the structural channel in step 3.2.1 i Embedded representation Z SR , For the node v of the feature channel in step 3.2.2 i Embedded representation Z FR ; Step 3.6.3: Compare the loss using the MLP decoder for the feature channel and the MLP decoder for the structure channel as described in Step 3.5.

2. Supervise the reconstruction process and use contrastive loss alignment. and Calculate the contrast loss of the decoder in For decoder reconstruction loss, For the reconstructed features of the feature channels and structural channels in step 3.5.2, (·,·) is the vector inner product, n is the total number of nodes, and τ is the adjustment coefficient; Step 3.6.4: Calculate the total loss function using weighted summation. Where λ1 is the encoder loss The weighting coefficients, where λ² is the decoder loss. Weighting coefficients; Step 3.6: Update the parameters of the graph completion learning model for the power system using the gradient backpropagation algorithm until convergence.

6. The power system target selection method under incomplete information conditions according to claim 4, characterized in that, Step 4: For the target power system, perform structural completion, node feature completion, and line feature completion based on the trained power system-oriented graph completion learning model. Calculate the confidence scores for structural completion, feature completion, line completion, and overall completion, where: (1) Confidence of structural completion The confidence level of structural completion reflects the reliability of topology completion, and the computation node v i The average connection probability with all other nodes; a higher value indicates more reliable topology completion. Where A comp ∈R N×N This is the completed probabilistic adjacency matrix. Represents node v i and v j The connection probability, where N is the total number of nodes; The node embedding representation Z of the structural channel output by the structural encoder (GCN) in step 3.2.1 SR Generate the completed node connection probabilities using MLP: Where θ is the Sigmoid function, and its output value is... Indicates the probability of node connection; (2) Confidence of node feature completion The confidence score of node feature completion is used to evaluate the quality of node feature repair. Based on the absolute error between the known node features and the completed node features, it is normalized and then mapped by tanh. Finally, it is converted into the confidence score of node feature completion by 1-tanh(·). in X is the completed node feature matrix. known Given the known node feature values, M∈{0,1} N×D The missing mask matrix is ​​denoted by ∈, where ∈ is the smoothing coefficient and D is the number of feature dimensions. (3) Confidence of line feature completion The confidence level of line feature completion is calculated by weighting the connection probability of the two-end nodes and the verification results of physical constraints. in Let S be the node connection probability of nodes s and t at both ends of line e in the completed adjacency matrix. For voltage effectiveness, For the validity of the number of parallel connections, For load effectiveness, α l =0.6, β l =0.4 is the weighting coefficient; (4) Overall confidence level C i : Calculate the line contribution: E i For node v i The set of connected lines, |E i | Indicates the number of associated lines; Determine the dynamic weighting coefficients: The basic weight λ1 = 0.7, and the attenuation coefficient λ2 = 0.1; Final overall confidence level: Where α s =0.6 is the structural confidence weight, α f =0.4 is the feature confidence weight.

7. The power system target selection method under incomplete information conditions according to claim 4, characterized in that, Step 5: Calculate the target importance of nodes to support the selection of important nodes in power systems under conditions of incomplete information. The specific method is as follows: Step 5.1: Select the target importance index, including the load impact index ΔP. ′ L Node degree index K ′ i Betweenness centrality index B ′ i The target's repair coefficient R ′ i Four categories of benefit indicators, and the structural information missingness indicator S miss and feature information missingness index F miss Two types of cost indicators are ranked by their importance using the TOPSIS method to support the selection of power system objectives under conditions of incomplete information. (1) Load impact index The load impact index assesses the importance of a target by evaluating the amount of load shedding directly or indirectly caused by damage to each target. Where M is the total equivalent load, N is the total number of power plants, and ΔP L ΔP represents the total load removed from the system after the target is damaged. 1,j Let P be the power flowing from the 1st node to the jth node. G,i Let P represent the power output of the i-th power plant. d,j Let ΔP represent the total load of the j-th equivalent node. d,j This represents the amount of load that needs to be removed from the j-th node when a certain target is damaged; If load data is missing, adjust the load impact indicators. Complete the confidence scores for node features; (2) Node degree index The node degree index assesses the importance of each target in the power system by analyzing the degree of each node, and uses an adjacency matrix to describe the relationships between nodes in the power system. If the structural adjacency matrix of the completed power system network is Then we know the degree K of the i-th node in the network. i satisfy: The relationship between the degree of a node in a power system network and the number of network edges X is as follows: If node v i The structure has high missing values ​​and low confidence, meaning its actual number of connections may be underestimated. Therefore, adjustments should be made to the node degree index. To complete the confidence level of the structure; (3) Betweenness centrality of nodes The betweenness centrality of a node is the proportion of the number of paths passing through that node in all shortest paths in a power system network, reflecting the importance of each power target in power transmission. node v i Betweenness centrality index B i Defined as: Where n jk Represents node v j v h The number of paths between them; n jh(i) Represents node v j v h The path passes through node v i The number of; If features or structures are missing, the betweenness centrality index is corrected, C i The overall confidence level calculated in step 4; B ′ i =B i ×C i (4) Target repair coefficient index The target repair coefficient index is evaluated from four perspectives: the region where the power target is located, the type of power target, the rated capacity of the power target, and the difficulty coefficient of the target's maintenance, using a subjective weighting method. Target repair coefficient R i The quantitative calculation formula is as follows: Where a1, a2, and a3 represent the weights of different indicators, the subscript i is used to distinguish different targets, and the subscript j represents different types of power targets: j=1 for power plants, j=2 for substations, and j=3 for power lines; o i This represents the material supply rate for repair, where K rk To analyze the number of dispatch range divisions within the power supply area of ​​this power grid, n k Indicates the k-th element to which the power target belongs. rk Total population of the region; p i U represents the cost of repair. i U represents the rated capacity of the i-th target. j,max This indicates the maximum capacity of the third type of target; L i L represents the distance of the i-th target from the scheduling center. max q represents the target farthest from the dispatch center within the dispatch area. j (j=1,2,3) represents the target repair coefficients for different types; If the node features have a high degree of missing information and low confidence, the uncertainty of the repair time increases, and the repair coefficient needs to be corrected. Where γ R C is the penalty coefficient, reflecting the sensitivity of missing node information to repair time; i For the overall confidence level calculated in step 4, I miss Overall information missingness α I β I These are the weighting coefficients. The degree of missing structural information. The degree of missing feature information; (5) Structural information missingness index: Structural information missing degree S miss Defined as the proportion of missing topological connections among nodes, reflecting incomplete structural information; (6) Feature information missingness index The feature information missingness index is the proportion of missing node feature fields, reflecting the incompleteness of attribute information; Feature information missing degree F miss definition: Among them, X i To complete the features of the preceding node, To complete the node features; Step 5.2: Use the TOPSIS method to rank the importance of power targets and obtain the final target importance ranking results for power system nodes; Step 5.2.1: Construct the original decision matrix D = [y ij ] n×6 n represents the target number of electricity consumption, with each row corresponding to 6 indicators, and m = 6, including the load impact indicator ΔP. ′ L Node degree index K ′ i Betweenness centrality index B ′ i and the target repair coefficient R ′ i Four categories of benefit indicators, and the structural information missingness indicator S miss and feature information missingness index F miss Two types of cost indicators; Normalize matrix D to obtain normalized matrix Z. The normalization formula is as follows: Step 5.2.2: Assign weights w to each attribute using the entropy weight method. j The weighted normalized decision matrix is ​​constructed as shown in the following equation; Where x ij =w j ×z ij w represents the weighted, normalized attribute value. j The weight of attribute j; Step 5.2.3: Calculate the ideal solution x according to the following formula. + and negative ideal solution x - ; Step 5.2.4: Calculate the power system node v i To the ideal solution x + and negative ideal solution x - The distance, using the Euclidean norm as the measure of distance, is shown in the following formulas; Step 5.2.5: Calculate the power system node v i Relative proximity R to the ideal solution i As shown in the following formula; R i The larger the value, the greater the importance of the goal, and vice versa, according to R. i The magnitude of the value can be used to rank the importance of nodes in the power system.

8. A power system target selection system under conditions of incomplete information, characterized in that, The method for selecting power system targets under incomplete information conditions as described in any one of claims 1-7 is implemented to achieve power system target selection under incomplete information conditions. Steps 1 to 5 are executed separately in five modules.

9. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, it implements the power system target selection method under incomplete information conditions as described in any one of claims 1-7, thereby realizing power system target selection under incomplete information conditions.

10. A computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, it implements the power system target selection method under incomplete information conditions as described in any one of claims 1-7, thereby realizing power system target selection under incomplete information conditions.

Images