Power distribution network robust state estimation method based on multi-modal graph neural network

Abstract

The invention belongs to the technical field of power distribution network situation awareness, and discloses a power distribution network robust state estimation method based on a multi-modal graph neural network, and the method comprises the steps: obtaining the real measurement data and historical operation information of a power distribution network; carrying out multi-modal graph structure fusion modeling; performing pseudo measurement generation mechanism and joint optimization; optimizing a robust graph neural network; and constructing a distributed state estimation architecture and a boundary cooperation mechanism. According to the power distribution network robust state estimation method based on the multi-modal graph neural network, substantive breakthroughs are achieved in four dimensions of improving the state estimation precision, enhancing the abnormal robustness, optimizing the reconstruction effect of a low observable area and supporting distributed deployment; the application bottleneck of a traditional estimation method under the conditions of data scarcity, dynamic disturbance and structure dispersion is solved, technical support is provided for constructing a high-reliability and high-intelligence power distribution network state sensing system, and the method has remarkable practical value and popularization potential.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of power distribution network situation awareness technology, and particularly relates to a power distribution network robust state estimation method based on a multi-modal graph neural network. BACKGROUND

[0002] With the large number of distributed energy access, the rapid development of electric vehicles and the promotion of user-side load response, the structure of modern power distribution network is increasingly complex, and the operating state changes frequently, which puts forward higher requirements for real-time perception and accurate estimation of its operating state. As an important basis for realizing power dispatching decision, operation control and fault warning, power distribution network state estimation occupies a core position in the power distribution automation system.

[0003] However, the traditional power distribution network state estimation method is usually based on mathematical models such as least square estimation and extended Kalman filter. This kind of method highly depends on complete, continuous and accurate measurement data input, and needs the system to have good observability. However, in the actual power distribution network, due to the low voltage level, the limited deployment of measurement equipment, and the shortage of communication resources, there are generally sparse measurement points, incomplete information and even long-term missing situations, which makes it difficult for the traditional method to accurately estimate key state parameters such as voltage and current, and further affects the safe operation and control efficiency of the power grid.

[0004] In recent years, artificial intelligence, especially deep learning technology, has been gradually applied in power system state estimation. Among them, graph neural network has become an important research direction in the field of power distribution network state estimation because of its ability to model the topology structure of the power grid. The existing graph neural network method can learn and infer the node state based on the electrical structure, and improve the perception ability of the model to the local structure. However, this kind of method generally has the following problems:

[0005] Firstly, it only relies on single topology modeling, and cannot effectively fuse the multi-source heterogeneous information relationship between the electrical connection structure and the measurement data, and cannot fully excavate the spatial correlation and information redundancy between the measurement data;

[0006] Secondly, when facing measurement data missing, mutation, abnormal interference or communication anomaly, the model has poor robustness and is prone to estimation deviation or failure;

[0007] Thirdly, the existing model is mostly a centralized processing framework, which is difficult to adapt to the development trend of future distributed energy access and micro-grid edge autonomous control, and has insufficient support for deployment efficiency and privacy protection.

[0008] Therefore, there is an urgent need for a new state estimation method that can simultaneously fuse electrical structure and measurement information, has data robustness, and supports distributed deployment, to adapt to the new situation of modern power distribution network operation and improve intelligent perception and autonomous control ability. SUMMARY

[0009] The application aims to provide a power distribution network robust state estimation method based on a multi-modal graph neural network, which closely surrounds the four core problems of "sparse measurement points, serious data missing, frequent abnormal disturbances and complex distributed operation" faced by the current power distribution network in actual operation. By systematically constructing a multi-graph model that integrates electrical structure and measurement information, introducing a pseudo-measurement generation and robust learning mechanism, and combining a distributed graph neural network framework design, the adaptability, accuracy and deployment efficiency of the state estimation system are significantly improved.

[0010] To achieve the above-mentioned purpose, the application provides a power distribution network robust state estimation method based on a multi-modal graph neural network, comprising the following steps:

[0011] S1, obtaining real measurement data and historical operation information of the power distribution network;

[0012] S2, multi-modal graph structure fusion modeling;

[0013] S3, pseudo-measurement generation mechanism and joint optimization;

[0014] S4, robust graph neural network optimization;

[0015] S5, constructing a distributed state estimation architecture and boundary coordination mechanism.

[0016] Preferably, S2 is specifically:

[0017] S21, state estimation problem modeling and multi-graph structure construction;

[0018] S22, constructing a multi-modal graph neural network and state estimation.

[0019] Preferably, S21 is specifically:

[0020] S211, constructing an electrical topology graph and a measurement point graph;

[0021] The electrical topology graph is as follows:

[0022] G e =(V,E e );

[0023] Wherein, G e is an electrical topology graph; V is a node set; E e is an electrical connection edge, and each edge has impedance; A e is an adjacency matrix; A e matrix is specifically represented as:

[0024]

[0025] Where i and j are node numbers respectively; Z ij is the impedance between node i and j;

[0026] The measurement point graph is shown as follows:

[0027] G m = (V m ,E m );

[0028] Where G m is the measurement point graph; m is the measurement device number; V m is the measurement device set; E m is the edge set between nodes, and the edge exists under the conditions of statistical correlation and geographical proximity. If the measurement value time series correlation Corr(X i ,X j )>δ, the edge exists, where δ is the correlation threshold; A m is the adjacency matrix, which is specifically represented as:

[0029] (A m ) ij =NMI(X i ,X j );

[0030] Where NMI(X i ,X j ) is the normalized mutual information of node i and j; X i and X j are the historical measurement time series data of i and j respectively.

[0031] S212, construct the heterogeneous mapping function between the measurement point and the node, as shown below:

[0032] f:V m →V;

[0033] S213, initialize the feature vector of each node, as shown below:

[0034]

[0035] Where P i , Q i are power values; is the node voltage amplitude measurement value; is the node voltage phase angle measurement value; type i is the node type code; flag i is the judgment of whether it is a measurement point.

[0036] Preferably, S22 is specifically:

[0037] S221, respectively extracting features of the electrical graph G e and the measurement graph G m , obtaining and wherein is the node embedding extracted from the electrical graph G e ; is the node embedding extracted from the measurement graph G m ;

[0038] S222, weighted integration is performed through a gating mechanism to obtain fused information, as follows:

[0039]

[0040] wherein, is the final fused representation of node i; Z i is the gating coefficient, specifically represented as: wherein, W z , b z are trainable parameters for fusion weights;

[0041] S223, the node state estimation result is obtained through a state estimation regressor, as follows:

[0042]

[0043] wherein, is the voltage amplitude estimation value of node i; is the phase angle estimation value of node i; MLP is a multilayer perceptron, and the output is a continuous value regression.

[0044] Preferably, S3 is specifically:

[0045] S31, a pseudo-measurement generation mechanism based on a CNN-BiLSTM-Attention neural network is introduced to complete the state information of unobserved nodes;

[0046] S311, local features of node data are captured through a convolutional neural network, as follows:

[0047]

[0048] wherein, is the feature representation of node i at time t; is the input data of node i at time t;

[0049] S312, long-term dependencies of time series data are processed through a bidirectional long short-term memory network, as follows:

[0050]

[0051] wherein, is the timing feature of node i at time t; is the output of the last time step;

[0052] S313, calculate the attention weight through the attention mechanism, optimize the feature selection, as follows:

[0053]

[0054] wherein, α ij is the attention weight between node i and node j; a is a learnable attention parameter; N i is the neighbor node set of node i; || is the vector splicing operation; u is a temporary variable when summing up all neighbors of node i; T is the transpose operation of the vector.

[0055] S32, through the state estimation model, the pseudo-measurement and the real measurement data are jointly optimized, wherein the joint optimization function L total(1) , as follows:

[0056] L total(1) = L SE + λ1L PM + λ2L phys ;

[0057] wherein, λ1, λ2 are hyperparameters for adjusting the weight of each loss; L SE is the state estimation error, specifically represented as: wherein, is the estimated voltage of node i; v i is the real voltage of node i; N is the number of nodes in the power grid; L PM is the pseudo-measurement generation error, specifically represented as: wherein, is the generated pseudo-measurement value; z i is the real measurement value; L phys is the physical consistency loss, specifically represented as: wherein, is the active power estimation value of node i output by the GNN; is the power calculated according to the estimated voltage and the phase angle .

[0058] Preferably, S4 is specifically:

[0059] S41, introduce the Dropout mechanism to simulate data missing; use the Gated Recurrent Unit GRU to model the historical measurement sequence, the specific content is as follows:

[0060] S411, Dropout mechanism, as follows:

[0061]

[0062] wherein, represents the screened features, represents the input features of node i at time t; is the corresponding Dropout mask, the elements of are 0 or 1, representing whether the feature is shielded;

[0063] S412, using a gated recurrent unit GRU to model the historical measurement sequence and handle missing data in the time series, specifically represented as:

[0064]

[0065] wherein, is the time series feature of node i at time t; is the time series feature of node at time t-1; is the input feature processed by the Dropout mask;

[0066] S42, introducing a credibility factor and an anomaly-aware attention mechanism to deal with the interference of abnormal values, as follows:

[0067] S421, introducing a credibility factor c i to adjust the information propagation strength, as follows:

[0068] c i =σ(W c ·h i );

[0069] wherein, σ is an activation function; W c is a learned weight matrix; h i is the feature representation of the node;

[0070] The credibility factor dynamically adjusts the attention weight between nodes, suppressing the influence of low confidence nodes;

[0071] S422, combining the attention mechanism to calculate the attention weight between nodes, as follows:

[0072]

[0073] S43, graph structure robustness constraint, by introducing a graph smoothing regularization term in the loss function, to promote the estimated values between adjacent nodes to be consistent, wherein the graph smoothing regularization term is as follows:

[0074]

[0075] wherein, L smooth is the graph smoothing regularizer; E is the edge set in the graph; is the estimated voltage of node i; is the estimated voltage of node j; w ij is the weight of edge, usually defined according to the electrical connectivity or other characteristics between nodes.

[0076] S44, the combination mechanism and comprehensive optimization, the final loss function, as follows:

[0077] L total(2) = L SE + λ3L smooth + λ4L confidence + λ5L dropout ;

[0078] wherein, L total(2) is the final loss function; L confidence is the loss based on the credibility factor; L dropout is the loss introduced by the Dropout mechanism; λ3, λ4, λ5 are the importance hyperparameters of adjusting the importance of each part of the loss.

[0079] Preferably, S5 is specifically:

[0080] S51, region division;

[0081] S52, the boundary node introduces a state sharing mechanism to improve state consistency and synergy;

[0082] S53, cross-region consistency loss.

[0083] Preferably, S51 is specifically:

[0084] S511, the overall power distribution network graph structure is divided into K non-overlapping subgraphs, as follows:

[0085]

[0086] Satisfy the following relationship:

[0087]

[0088] wherein, G k = (V k , E k ) is the kth subgraph; V k is the node set of the kth subgraph; E k is the edge set of the kth subgraph; k is the subgraph serial number.

[0089] S512, based on the electrical distance of the region clustering division strategy, improve the physical rationality and K region autonomy ability of the division, the specific content is as follows:

[0090] Firstly, the electrical distance between nodes i and j is defined as follows:

[0091]

[0092] Wherein, d ij is the electrical distance between nodes i and j; the smaller the electrical distance represents the stronger the electrical coupling between the two nodes; Z ii is the impedance of node i; Z ij is the impedance of node i and j; Z jj is the impedance of node j.

[0093] Secondly, the electrical distance matrix is constructed as follows:

[0094] D=[d ij ];

[0095] Finally, through the spectral clustering algorithm, the region division is realized, and the optimization target is as follows:

[0096]

[0097] Preferably, S52 is specifically:

[0098] Let node i∈V k ∩V1 is the region boundary node, and the final fusion state of the node is as follows:

[0099]

[0100] Wherein, is the final fusion state of the node; is the hidden state output by the local GNN module; is the state received from the adjacent region; Fuse(·) is the fusion function.

[0101] Preferably, S53 is specifically:

[0102] S531, introduce the cross-region state consistency constraint term, ensure the cross-region consistency of the estimated result in physical meaning, as follows:

[0103]

[0104] Wherein, L cross is the cross-region state consistency constraint term; are the state vectors estimated by the boundary nodes in the respective region GNN; Border is a set composed of all cross-region boundary node pairs;

[0105] S532, obtaining an optimal state estimation result by final joint optimization target L total(3) , as follows:

[0106]

[0107] wherein, is the local estimation loss within the region k, and λ is the cross-region consistency regularization term weight.

[0108] Therefore, the power distribution network robust state estimation method based on the multi-modal graph neural network has the following beneficial effects:

[0109] (1) In terms of state estimation accuracy, the application constructs a double graph structure input system including an electrical topology graph and a measurement point information graph, which respectively describes the functional and statistical correlation between the physical connection relationship of the power distribution network and the measurement points. By introducing a graph attention mechanism and a gating fusion mechanism, dynamic extraction and weighted fusion of information in the two types of graph structures are realized, thereby improving the accuracy of node state estimation and the adaptability of the model to multi-source information.

[0110] (2) The application proposes a pseudo-measurement point generation module based on historical data driving to solve the problem of missing state information due to the lack of measurement devices at some nodes in the actual power distribution network. The state information of unobserved nodes is completed by using a generative model. The pseudo-measurement information is used as auxiliary input and is jointly optimized with the state estimation module for training, which can significantly improve the state recovery capability of the low observability area and ensure the interpretability and electrical consistency of the estimation result through the physical constraint regularization term.

[0111] (3) In terms of abnormal disturbance and communication interruption, in order to enhance the ability of the model to deal with non-ideal situations such as measurement data missing, communication interruption, and abnormal value disturbance in actual operation, the application introduces a Dropout mechanism, a node confidence adjustment factor, and an abnormal perception attention mechanism in the graph neural network. At the same time, combined with the graph structure disturbance robustness regularization term, the stability and generalization ability of the model under complex working conditions are improved.

[0112] (4) In terms of system deployment capability, in order to meet the demand for regional autonomy and edge deployment under the background of distributed energy access, the application designs a distributed graph neural network estimation architecture that supports regional division and boundary communication. The graph neural network module runs independently in each region, the boundary nodes update the state through an information sharing mechanism, and the cross-region consistency constraint is introduced to improve the overall estimation coordination. This architecture has good engineering adaptability and scalability, and is suitable for various practical application scenarios of power distribution networks.

[0113] (5) The power distribution network state estimation method provided by the application can provide stable, accurate and deployable state estimation results under the conditions of incomplete information, strong structure dynamics and complex operation environment, and has good engineering practical value and technical popularization prospect.

[0114] The technical solutions of the application will be further described in detail below with reference to the drawings and examples. BRIEF DESCRIPTION OF DRAWINGS

[0115] Figure 1 is a flowchart of a power distribution network robust state estimation method based on a multi-modal graph neural network;

[0116] Figure 2 is a method flowchart of a power distribution network robust state estimation method based on a multi-modal graph neural network;

[0117] Figure 3 is a convolutional neural network CNN flowchart in the embodiment of the application;

[0118] Figure 4 is a bidirectional long short-term memory network BiLSTM neural network schematic diagram in the embodiment of the application;

[0119] Figure 5 is an attention mechanism Attention unit structure diagram in the embodiment of the application. DETAILED DESCRIPTION

[0120] The technical solutions of the application will be further described in detail below with reference to the drawings and examples.

[0121] EMBODIMENT

[0122] As shown in Figures 1-2 , a power distribution network robust state estimation method based on a multi-modal graph neural network comprises the following steps:

[0123] S1, obtaining simulation data and historical data of the power distribution network.

[0124] S2, multi-modal graph structure fusion modeling.

[0125] S21, state estimation problem modeling and multi-graph structure construction.

[0126] S211, constructing an electrical topology graph and a measurement point graph.

[0127] The electrical topology graph is as follows:

[0128] G e =(V,E e );

[0129] Wherein, G eis an electrical topology; V is a set of nodes, such as busbars, distribution transformers and other physical nodes; E e is a branch, such as a feeder, a cable and other electrical connection edges, and each edge has an impedance weight; A e is an adjacency matrix; A e is specifically represented by the matrix:

[0130]

[0131] where i and j are node numbers; Z ij is the impedance between nodes i and j.

[0132] A measurement point graph is as follows:

[0133] G m = (V m , E m );

[0134] where G m is a measurement point graph; m is a measurement device number; V m is a smart meter, a PMU, an FTU and other measurement devices; E m is a set of edges between nodes, and the existence of an edge is based on statistical correlation and geographical proximity. If the correlation Corr(X i , X j )>δ, the edge exists, where δ is a correlation threshold; A m is an adjacency matrix, which is a mutual information value.

[0135] To effectively mine the statistical correlation between each measurement node of the power distribution network, the embodiment introduces a mutual information coefficient (Mutual Information, MI) as the weight of the adjacency matrix in the graph neural network. Mutual information is a nonlinear index that measures the degree of information sharing between random variables, which can capture linear and nonlinear correlation relationships and is suitable for modeling the statistical dependence between continuous variables such as voltage and current.

[0136] Specifically, let X i , X j represent the historical measurement time series of nodes i and j, respectively, and the mutual information is defined as:

[0137]

[0138] where I(X i ; X j ) is the mutual information value; p(x i , x j ) is the joint probability density function; and p(x i ) and p(x j ) are the marginal probability density functions.

[0139] Due to the difficulty of analytical calculation of the integral form, a non-parametric estimation method based on k-nearest neighbors is used to estimate the joint probability density and marginal probability density, so as to obtain the mutual information value.

[0140] In order to facilitate the use in the graph neural network and enhance the comparability of the mutual information value between different nodes, the normalized mutual information (NMI) is further introduced as the final edge weight calculation standard, as follows:

[0141]

[0142] where H(X i ) and H(X j ) represent the information entropy of X i and X j , respectively.

[0143] The normalized mutual information value is limited in the interval [0, 1], which has good numerical stability and physical interpretability.

[0144] Finally, the adjacency matrix A m is constructed based on the NMI value between all node pairs, which is specifically represented as:

[0145] (A m ) ij =NMI(X i ,X j );

[0146] where NMI(X i ,X j ) is the normalized mutual information calculated based on the historical measurement data of nodes i and j; X i and X j represent the historical measurement time series data of nodes i and j, respectively.

[0147] The adjacency matrix not only reflects the statistical dependence relationship between nodes in the physical structure of the power distribution network, but also provides a data-driven structural prior for subsequent spatial information modeling of the graph neural network.

[0148] S212, construct a heterogeneous mapping function between the measurement points and the nodes, as follows:

[0149] f:V m →V;

[0150] S213, initialize the feature vector of each node, as follows:

[0151]

[0152] wherein P i , Q i are power values; is a node voltage amplitude measurement value; is a node voltage phase angle measurement value;type i is a node type code;flag i is a judgment of whether it is a measurement point.

[0153] S22, constructing a multi-modal graph neural network and state estimation.

[0154] S221, using a parallel graph neural network to respectively perform feature extraction on the electrical graph G e and the measurement graph G m .

[0155] Through graph convolution, the information on the two subgraphs is respectively extracted as and wherein is the node embedding extracted from the electrical graph G e , is the node embedding extracted from the measurement graph G m .

[0156] Taking a graph attention network as an example, the (l+1)th layer feature of node i on the electrical graph G e is as follows:

[0157]

[0158] wherein, is the graph embedding of node i at the (l+1)th layer; σ(·) is an activation function; is the neighbor set of node i in the graph G e ; W e is a linear transformation matrix of the electrical graph; is the hidden representation of node j at the lth layer; is the attention weight of node j to node i.

[0159] Using the same structure on the measurement graph G m , the (l+1)th layer feature representation of node i on the measurement graph G e is obtained , which represents the embedding obtained from the perspective of the measurement point.

[0160] S222, weighted integration is performed through a gating mechanism to obtain the fused information, as follows:

[0161]

[0162] wherein, is the final fused representation of node i; Z iis a gating coefficient, which is specifically expressed as: wherein, W z , b z is a trainable parameter for fusion weight.

[0163] S223, through the state estimation regressor, the node state estimation result is obtained, as follows:

[0164]

[0165] wherein, is the voltage amplitude estimation value of node i; is the phase angle estimation value of node i; MLP is a multilayer perceptron, and the output is a continuous value regression.

[0166] In a low observable scenario, some nodes have no real-time measurement value. To alleviate this problem, a pseudo measurement point generation module is introduced to generate estimated values through historical measurement sequences and electrical structure information, thereby improving the overall state estimation accuracy.

[0167] S3, pseudo measurement generation mechanism and joint optimization.

[0168] S31, a pseudo measurement generation mechanism based on CNN-BiLSTM-Attention neural network is introduced to complete the state information of unobserved nodes.

[0169] In a distribution network, in order to obtain more node data information, a neural network is used to add pseudo measurements to the distribution network, and a pseudo measurement model is established to help better understand the state of the power grid, thereby realizing more effective monitoring and management.

[0170] The pseudo measurement generation mechanism is based on CNN-BiLSTM-Attention neural network for node data prediction. The CNN-BiLSTM-Attention neural network combines convolutional neural network CNN, bidirectional long short-term memory network BiLSTM and attention mechanism. The combination of these three technologies makes the neural network have stronger feature extraction ability, higher prediction accuracy and better time series modeling ability.

[0171] First, the convolutional neural network is a neural network with local perception and weight sharing characteristics, which can extract features while preserving spatial information. In the prediction of distribution network node data, CNN can be used to capture the local features of node data, thereby improving the prediction performance of the model, as shown in Figure 3 .

[0172] Based on this, S311, through the convolutional neural network, the local features of node data are captured, as follows:

[0173]

[0174] wherein, is the feature representation of node i at time t; is the input data of node i at time t.

[0175] Secondly, the bidirectional long short-term memory network is a kind of recurrent neural network capable of capturing long-term dependencies in time series data. Unlike traditional LSTM, BiLSTM considers both past and future information at each time step. This bidirectional structure enables BiLSTM to better capture the time series relationship in the data when processing the node data of the power distribution network, thereby improving the accuracy of the prediction.

[0176] The BiLSTM neural network is composed of two LSTM neural networks with the same input but different information transmission directions. One is trained forward, and the other is trained backward, and the final output result is determined by the LSTM neural networks in both forward and backward directions, as shown in Figure 4 .

[0177] h t =f(w1X t +w2h t-1 );

[0178] h' t =f(w3X t +w5h' t+1 );

[0179] O t =g(w4h t +w6h' t );

[0180] where X t is the input at time t, w1, w2, w3, w4, w5, w6 are the corresponding weight matrices, h t-1 is the output of the previous time, h t is the output of the forward layer at time t, h' t+1 is the output of the next time, h' t is the output of the reverse layer at time t, O t is the output at time t.

[0181] Based on the above, in the present embodiment, S312, the long-term dependence relationship of the time series data is processed by the bidirectional long short-term memory network, as follows:

[0182]

[0183] wherein, is the time series feature of node i at time t; is the output of the previous time.

[0184] Finally, attention mechanism is a technique that can assign different weights to different input data. In the CNN-BiLSTM-Attention neural network, attention mechanism can help the model identify key time steps and features, so as to better focus on information useful for the prediction task. This can further improve the performance of the model in processing the node data prediction task of the power distribution network.

[0185] Attention originated from the simulation of human brain attention characteristics, and this method was first applied in the field of image processing. In the field of deep learning, the Attention mechanism assigns different weights to different features, giving larger weights to key content and smaller weights to other content. Through differentiated weight allocation, the efficiency of information processing can be improved. The structure of the Attention unit is as shown in Figure 5 .

[0186] The attention state transition process is as follows:

[0187]

[0188] wherein: α ti is the hidden layer output value h t of the BiLSTM for the current input; y1, y2, y3, … y t is the input sequence; h1, h2, h3, … h t is the hidden layer state value corresponding to the input sequence y1, y2, y3, … y t , that is, h t is the hidden layer state value corresponding to the input y t ; h' t is the final feature vector; V, W, U, b are learning parameters of the model, which are updated constantly during the model training process.

[0189] By collecting equipment measurement data, generator output data and other factors in the power distribution network as the input of the neural network, the output is the data of the important nodes or the nodes adjacent to the important nodes without monitoring and data acquisition system, phasor measurement unit and other observers.

[0190] Based on this, S313 calculates the attention weight through the attention mechanism and optimizes the feature selection, as shown below:

[0191]

[0192] wherein: α ij is the attention weight between node i and node j; a is a learnable attention parameter; N iis the set of neighbor nodes of node i; u is all neighbors of node i; T is the transpose operation.

[0193] S32, jointly optimizing the pseudo-measurement and the real measurement data through a state estimation model, wherein the joint optimization function L total(1) As follows:

[0194] L total(1) = L SE + λ1L PM + λ2L phys ;

[0195] Wherein, λ1, λ2 are hyperparameters for adjusting the weight of each loss, controlling the contribution of each loss to the final goal; L SE is the state estimation error, specifically represented as: Wherein, is the estimated voltage of node i; v i is the real voltage of node i; N is the number of nodes in the power grid; L PM is the pseudo-measurement generation error, specifically represented as: Wherein, is the generated pseudo-measurement value; z i is the real measurement value; L phys is the physical consistency loss, specifically represented as: Wherein, is the active power estimate value of node i output by the GNN; is the power calculated according to the estimated voltage and the phase angle .

[0196] S4, robust graph neural network optimization.

[0197] S41, introduce Dropout mechanism to simulate data missing; use gated recurrent unit GRU to model time series features.

[0198] S411, Dropout mechanism, as follows:

[0199]

[0200] Wherein, represents the filtered features, represents the input features of node i at time t; is the corresponding Dropout mask, the elements of which are 0 or 1, representing whether the feature is shielded.

[0201] S412, in order to further process the missing data in the time sequence, a gated recurrent unit (GRU) is used to model the time sequence feature. The GRU can effectively capture the time dependence and process the historical information of the node state, which is specifically represented as:

[0202]

[0203] wherein, is the time sequence feature of node i at time t; is the time sequence feature of node i at time t-1; is the input feature processed by the Dropout mask.

[0204] S42, a credibility factor and an abnormal perception attention mechanism are introduced to deal with the interference of abnormal values. By evaluating the confidence of each node, the model can suppress the influence of low-confidence nodes during state estimation and avoid the interference of abnormal values on the final result.

[0205] S421, a credibility factor c i is introduced to adjust the information propagation strength, as shown below:

[0206] c i =σ(W c ·h i );

[0207] wherein, σ is an activation function; W c is a learned weight matrix; h i is the feature representation of the node.

[0208] The credibility factor can dynamically adjust the attention weight between nodes and suppress the influence of low-confidence nodes.

[0209] S422, combined with the attention mechanism, the attention weight between nodes is calculated, as shown below:

[0210]

[0211] S43, a graph smoothing regular term is introduced to promote the consistency of the estimated values between adjacent nodes and improve the stability of the graph neural network in complex topologies and disturbance environments, wherein the graph smoothing regular term is as follows:

[0212]

[0213] wherein, L smooth is the graph smoothing regular term; E is the edge set in the graph; is the estimated voltage of node i; is the estimated voltage of node j; w ijis the weight of the edge, usually defined according to the electrical connectivity or other characteristics between nodes.

[0214] The above mechanism helps to reduce the fluctuation of state estimation values between adjacent nodes, and improve the robustness of the model to topology disturbance.

[0215] S44, the combination mechanism and comprehensive optimization, the final loss function, as follows:

[0216] L total(2) = L SE + λ3L smooth + λ4L confidence + λ5L dropout ;

[0217] Wherein, L total(2) is the final loss function; L confidence is the loss based on the credibility factor; L dropout is the loss introduced by the Dropout mechanism; λ3, λ4, λ5 are important parameters that adjust the importance of each part of the loss.

[0218] By jointly optimizing the above losses, the model can maintain high robustness and accuracy when dealing with missing data, outliers and disturbances.

[0219] To improve the scalability, deployment flexibility and parallel computing capability of the state estimation method in large-scale distribution network, the embodiment adopts a distributed state estimation architecture and a boundary coordination mechanism. This method fully considers the physical hierarchy and communication boundary constraints of the power grid structure, and realizes efficient deployment and accurate fusion in multi-region state estimation tasks through three mechanisms of region division, boundary state sharing and consistency optimization.

[0220] S5, constructing a distributed state estimation architecture and a boundary coordination mechanism.

[0221] S51, region division.

[0222] S511, divide the whole distribution network graph structure into K non-overlapping subgraphs according to certain criteria, as follows:

[0223]

[0224] Satisfy the following relationship:

[0225]

[0226] Wherein, G k = (V k , E k ) is the kth subgraph; V k is the node set of the kth subgraph; E kis the edge set of the kth subgraph; k is the subgraph sequence number.

[0227] S512, based on the electrical distance of the region clustering division strategy, improve the physical rationality and regional K domain autonomy ability of the division, the specific content is as follows:

[0228] First, define the electrical distance between nodes i and j, as follows:

[0229]

[0230] Where d ij is the electrical distance between nodes i and j; the smaller the electrical distance represents the stronger the electrical coupling between the two nodes; Z ii is the impedance of node i; Z ij is the mutual impedance of nodes i and j; Z jj is the impedance of node j.

[0231] Second, construct the electrical distance matrix, as follows:

[0232] D = [d ij ].

[0233] Finally, through the spectral clustering algorithm, realize the region division, the optimization goal, as follows:

[0234]

[0235] This way ensures that the coupling between nodes in each sub-region is strong and the structure is compact, which facilitates local GNN model learning and state perception. In addition, the region division can also be combined with geographical information, communication topology and measurement correlation and other factors to optimize the joint modeling to improve the cross-regional collaboration efficiency.

[0236] S52, the boundary node introduces the state sharing mechanism, improves the state consistency and collaboration.

[0237] Let node i∈V k ∩V1 is the region boundary node, and the final fusion state of the node is as follows:

[0238]

[0239] Where, is the final fusion state of the node; is the hidden state output by the local GNN module; is the state received from the adjacent region; Fuse(·) is a fusion function, which can be defined as weighted average, gating mechanism or attention mechanism, as follows:

[0240]

[0241] Where, Wg , b g are learnable parameters. This mechanism can adaptively adjust the weights of local and neighborhood information, and improve the expression ability of boundary state.

[0242] S53, cross-zone consistency loss.

[0243] S531, introduce cross-zone state consistency constraint term to ensure the cross-zone consistency of the estimated results in the physical sense, as follows:

[0244]

[0245] wherein, L cross is the cross-zone state consistency constraint term; are the state vectors estimated by the boundary nodes in their respective regional GNNs, such as voltage amplitude and phase angle; Border is a set composed of all cross-regional boundary node pairs.

[0246] S532, obtain the optimal state estimation result by optimizing the final joint optimization target, wherein the final joint optimization target L total(3) , as follows:

[0247]

[0248] wherein, is the local estimation loss within the region k, and λ is the cross-zone consistency regularization term weight.

[0249] By optimizing this loss function, the overall coordination and physical consistency of the whole network state estimation can be improved while maintaining the local accuracy.

[0250] Therefore, the power distribution network robust state estimation method based on the multi-modal graph neural network has the following beneficial effects:

[0251] (1) In terms of state estimation accuracy, the dual-graph structure effectively utilizes the statistical correlation information between the physical connection attributes of the power distribution network and the measurement points. Especially in low-observable areas, the gating fusion mechanism significantly enhances the perception and reconstruction ability of key node states. In cooperation with the pseudo-measurement point generation module, this method can still maintain the voltage estimation error below 2% level in the scene where the node monitoring missing rate exceeds 50%, which reflects strong model completion ability and accuracy.

[0252] (2) In terms of coping with abnormal disturbances and communication interruptions, the model introduces Dropout, confidence adjustment and abnormal perception attention mechanism during the training process, effectively improving the network's ability to identify and tolerate missing, noisy and abnormal data. In the test, when the input data contains 10% random noise or abnormal values, the estimation bias increases by no more than 0.5%, significantly better than the traditional least squares estimation and standard GNN model, verifying the robustness and practicality of the method under non-ideal working conditions.

[0253] (3) In terms of system deployment capability, the distributed estimation architecture designed in the application supports regional division of the distribution network and parallel deployment of GNN modules, especially for modern distribution systems with edge intelligent control requirements. The boundary node states are shared and consistent constraints are optimized among sub-graph models to ensure the overall estimation results of the system are consistent. Experiments show that this architecture can achieve fast parallel reasoning in large-scale multi-region distribution networks, with a system response time reduced by about 30%, and good scalability and deployment flexibility.

[0254] Finally, it should be noted that: the above examples are only used to illustrate the technical solutions of the present application and not to limit it, although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand: it can still modify or equivalently replace the technical solutions of the present application, and these modifications or equivalent replacements cannot make the modified technical solutions deviate from the spirit and scope of the technical solutions of the present application.

Claims

1. A robust state estimation method for distribution networks based on multimodal graphical neural networks, characterized in that, Includes the following steps: S1. Obtain real measurement data and historical operation information of the power distribution network; S2, Multimodal graph structure fusion to build two models; S3, Pseudo-measurement generation mechanism and joint optimization; S4, Robust Graph Neural Network Optimization; S5. Construct a distributed state estimation architecture and boundary coordination mechanism.

2. The robust state estimation method for distribution networks based on multimodal graphical neural networks according to claim 1, characterized in that, S2 specifically refers to: S21. Modeling of State Estimation Problems and Construction of Multi-Graph Structures; S22. Construct a multimodal graphical neural network and perform state estimation.

3. The robust state estimation method for distribution networks based on multimodal graphical neural networks according to claim 2, characterized in that, S21 specifically refers to: S211. Construct the electrical topology diagram and measurement point diagram, the specific contents of which are as follows; The electrical topology diagram is shown below: G e =(V,E e ); Among them, G e It is an electrical topology diagram; V is a set of nodes; E e These are electrical connection edges, and each edge has an impedance weight; A e It is an adjacency matrix; A e The matrix is ​​specifically represented as: Where i and j are the node numbers, respectively; Z ij It is the impedance between i and j; The measurement point map is shown below: G m =(V m ,E m ); Among them, G m This is a measurement point map; m is the measurement equipment number; V m It is a collection of measuring devices; E m It is the set of edges between nodes. The conditions for the existence of edges are statistical correlation and geographical proximity. If the time series correlation of the measured values ​​is Corr(X) i ,X j If A > δ, then connect the edges, where δ is the relevance threshold; m It is an adjacency matrix, specifically represented as: (A m ) ij =NMI(X i ,X j ); Among them, NMI(X) i ,X j X is the normalized mutual information between the measurement sequences of nodes i and j; i and X j These are the time series data for measuring nodes i and j, respectively. S212. Construct the heterogeneous mapping function between measurement points and nodes, as shown below: f:V m →V; S213. Initialize the feature vector for each node as follows: Among them, P i Q i It is the power value; It is the measured value of the node voltage amplitude; It is the measured value of the node voltage phase angle; type i It is a node type encoding; flag i It determines whether a point is a measurement point.

4. The robust state estimation method for distribution networks based on multimodal graphical neural networks according to claim 2, characterized in that, S22 specifically refers to: S221. Parallel graph neural networks are used to process electrical diagrams G respectively. e and measurement map G m Feature extraction is performed to obtain node embeddings. and in, From electrical diagram G e Extracted node feature embedding; From the measurement map G m Extracted node embedding; S222. Weighted integration is performed through a gating mechanism to obtain the fused information, as shown below: in, Z is the final fused representation of node i; i It is the gating factor, specifically expressed as: Among them, W z b z These are trainable parameters used for fusion weights; S223. The node state estimation results are obtained through the state estimation regressor, as shown below: in, It is the estimated voltage magnitude at node i; It is the phase angle estimate of node i; MLP is a multilayer perceptron, and its output is a continuous value regression.

5. The robust state estimation method for distribution networks based on multimodal graphical neural networks according to claim 1, characterized in that, S3 specifically refers to: S31. Introduce a pseudo-measurement generation mechanism based on CNN-BiLSTM-Attention neural network to complete the state information of unobserved nodes; S311. Using a convolutional neural network, local features of node data are captured, as shown below: in, It is the feature representation of node i at time t; It is the input data of node i at time t; S312. Long-term dependencies in time-series data are processed using a bidirectional long short-term memory network, as shown below: in, It is the temporal characteristic of node i at time t; It is the output from the previous moment; S313. Calculate attention weights and optimize feature selection through an attention mechanism, as shown below: Where, α ij is the attention weight between node i and node j; 'a' is the learnable attention parameter; N i It is the set of neighboring nodes of node i; [·||·] is the vector concatenation operation; u is the temporary variable when summing, and iterates through all neighbors u of node i; T is the transpose of the vector; S32. Using a state estimation model, jointly optimize the pseudo-measurement and real measurement data, where the joint optimization function L... total(1) As shown below: L total(1) =L SE +λ1L PM +λ2L phys ; Where λ1 and λ2 are hyperparameters that adjust the weights of each loss; L SE It is the state estimation error, specifically expressed as: in, It is the estimated voltage of node i; v i It is the actual voltage of node i; N is the number of nodes in the power grid; L PM This is a pseudo-measurement generation error, specifically expressed as: in, These are generated pseudo-measurements; z i It is the actual measured value; L phys It is the loss of physical consistency, specifically expressed as: in, It is the active power estimate of node i output by the GNN; It is based on the estimated voltage With phase angle Calculated power.

6. The robust state estimation method for distribution networks based on multimodal graphical neural networks according to claim 1, characterized in that, S4 specifically refers to: S41. Introducing the Dropout mechanism to simulate missing data; using a gated recurrent unit (GRU) to model historical measurement sequences, the specific details are as follows: S411, Dropout mechanism, as shown below: in, Indicates the filtered features, This represents the input features of node i at time t; It corresponds to the Dropoutmask. The element is 0 or 1, which indicates whether the feature is masked; S412. A gated recurrent unit (GRU) is used to model historical measurement sequences and handle missing data in the time series, specifically as follows: in, It is the temporal characteristic of node i at time t; It is the temporal characteristic of node i at time t-1; These are the input features after being processed by Dropout masking; S42. Introduce a credibility factor and an anomaly-aware attention mechanism to address the interference of outliers. Details are as follows: S421. Introduce the credibility factor c i Adjust the intensity of information dissemination as follows: c i =σ(W c ·h i ); Where σ is the activation function; W c It is the weight matrix for learning; h i It is a feature representation of a node; S422. Combining the attention mechanism, calculate the attention weights between nodes, as shown below: S43. Graph structure robustness constraint: By introducing a graph smoothing regularization term into the loss function, the estimated values ​​between adjacent nodes are kept consistent. The graph smoothing regularization term is as follows: Among them, L smooth E is the graph smoothing regularization term; E is the edge set in the graph; It is the estimated voltage of node i; It is the estimated voltage at node j; w ij It is the weight of the edge, which is usually defined based on the electrical connectivity or other characteristics between nodes; S44. Combination Mechanism and Comprehensive Optimization, the final loss function is shown below: L total(2) =L SE +λ3L smooth +λ4L confidence +λ5L dropout ; Among them, L total(2) It is the final loss function; L confidence It is a loss based on the credibility factor; L dropout λ3, λ4, and λ5 are the loss introduced by the Dropout mechanism; λ3, λ4, and λ5 are important hyperparameters for adjusting the loss of each part.

7. The robust state estimation method for distribution networks based on multimodal graphical neural networks according to claim 1, characterized in that, S5 specifically refers to: S51, Regional Division; S52. A state sharing mechanism is introduced at boundary nodes to improve state consistency and coordination; S53, Cross-regional consistency loss.

8. The robust state estimation method for distribution networks based on multimodal graphical neural networks according to claim 7, characterized in that: S51 specifically refers to: S511. Divide the overall power distribution network diagram structure into K non-overlapping sub-diagrams, as shown below: The following relationship must be satisfied: Among them, G k =(V k E k ) is the k-th subgraph; V k It is the set of nodes of the k-th subgraph; E k It is the edge set of the k-th subgraph; k is the subgraph index; S512. A region clustering strategy based on electrical distance improves the physical rationality of the partitioning and the autonomy of the K regions. The specific details are as follows: First, define the electrical distance between nodes i and j as follows: Where, d ij Z is the electrical distance between nodes i and j; ii Z is the impedance of node i; ij Z represents the impedance at nodes i and j; jj It is the impedance of node j; Secondly, construct the electrical distance matrix as follows: D=[d ij ]; Finally, the region division and optimization objective are achieved through spectral clustering algorithm, as shown below:

9. A robust state estimation method for distribution networks based on multimodal graphical neural networks according to claim 7, characterized in that, S52 specifically refers to: Let node i∈V k ∩V1 is the boundary node of the region. The final merged state of the node is shown below: in, This is the final fusion state of the nodes; It is the hidden state output by the local GNN module; It is the state received from the adjacent region; Fuse(·) is the fusion function.

10. A robust state estimation method for distribution networks based on multimodal graphical neural networks according to claim 7, characterized in that, S53 specifically refers to: S531. Introduce cross-region consistency loss to ensure that the estimation results are physically consistent across regions, as shown below: Among them, L cross This is a loss of consistency across regions; These are the state vectors estimated by the boundary nodes in their respective regional GNNs; Border is the set of all cross-regional boundary node pairs. S532, Through final joint optimization of objective L total(3) The optimal state estimation results are as follows: in, λ is the local estimated loss within region k, and λ is the weight of the cross-region consistency regularization term.

Images