Power system abnormal remote signaling detection method based on graph autoencoder model

Abstract

The application discloses a power system abnormal remote signaling detection method based on a graph autoencoder model, which comprises the following steps: collecting measurement data of a power system and performing pretreatment to obtain a graph data set with an abnormal label and divide the graph data set into a training set and a test set; inputting graph data in the training set into a graph autoencoder model for training; after the training is completed, an abnormal score distribution is statistically obtained based on normal edge samples in the training set, and a threshold is set as a subsequent judgment basis, and the threshold is selected with the accuracy and recall rate on the test set as the optimization target; in the test stage, graph data in the test set is inputted for edge feature reconstruction and abnormal scoring, and the set threshold is combined to judge the abnormality of the edge, and an initial abnormal edge detection result is outputted; the outputted abnormal edge detection result is inputted into a graph repair module, and a repaired edge structure and edge feature are outputted, the damaged remote signaling state in the power grid is recovered, and the integrity of the graph structure and the operation reliability of the power system are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power system abnormal remote signaling detection, and particularly to a power system abnormal remote signaling detection method based on a graph autoencoder model. BACKGROUND

[0002] With the rapid development of new power systems, the data scale and complexity required by power systems are rapidly increasing, which brings new challenges to information collection, processing and analysis. In order to ensure the safe and stable operation of the power system, the dispatching center of the power system needs to comprehensively understand the real-time operation state of the power system and be able to predict its operation trend. State estimation is a key technology in the situation awareness system, aiming to accurately perceive the operation state of the system. Accurate state estimation results play a crucial role in power system applications such as optimal power flow calculation, load forecasting, economic dispatching, etc. However, the accuracy of state estimation is highly dependent on the accuracy and timeliness of the collected data. In actual operation, due to switch (which can be referred to as remote signaling) state false positives (such as a closed switch being falsely reported as open), node measurement noise (such as abnormal fluctuations in voltage and current), and dynamic topology changes of the power grid (such as fault isolation or load switching), the monitoring system is prone to misjudgment, which in turn leads to incorrect dispatching decisions, and may even trigger a chain of failures. In addition, SCADA and other measurement devices may produce topology errors or bad measurement data in information collection due to device failure, data transmission communication anomalies, device aging, environmental interference, and sudden changes in system operation state. These abnormal data not only seriously affect the accuracy of state estimation, but also may lead to a decrease in the convergence performance of power system information processing, thereby increasing the risk of power grid operation. Therefore, effective power grid topology anomaly detection and timely correction of measurement data are of great practical significance for improving the reliability of state estimation results, ensuring data quality, and enhancing the safe and stable operation of the power system.

[0003] Information collection in power systems usually relies on SCADA systems, which collect data every 15 minutes, so the data presents obvious dynamic change characteristics. Traditional static graph modeling methods cannot fully reflect the dynamic change process of the operation state of the power system. In contrast, dynamic graph modeling can more accurately capture the evolution process of the topology structure and measurement data of the power grid over time, and is more in line with actual operation. However, due to the dynamic evolution characteristics of the nodes and edges of the dynamic graph, it brings significant challenges in data analysis, especially in anomaly detection. On the one hand, the nodes and edges in the dynamic graph may change frequently, and the abnormal features may be dynamic, local or global; on the other hand, this change increases the complexity of the data, making it difficult for traditional anomaly detection methods to effectively capture abnormal patterns in the dynamic process.

[0004] Anomaly detection, as a key task to ensure the safe operation of power systems and data integrity, aims to identify and locate abnormal events that significantly deviate from the normal mode in the data set. Timely and effective detection of anomalies is of great significance to improve the reliability of the system and reduce potential security risks. However, traditional anomaly detection methods usually rely on static feature analysis, making it difficult to effectively handle the dynamic characteristics of data and achieve efficient and accurate detection in high dynamic environments.

[0005] In recent years, dynamic graph anomaly detection methods based on deep learning have made significant progress and have been successfully applied in social networks, knowledge graphs, and network security, among other fields. These methods typically make full use of the inherent temporal characteristics and relationship structures of dynamic graphs, enabling effective identification of abnormal events in the data and thus improving the security and integrity of various network systems. For example, NetWalk proposes a node vector representation method based on deep autoencoders, combining clique embedding and reservoir sampling techniques to achieve efficient updating of dynamic graphs and effectively identifying structural anomalies in networks through clustering algorithms. AddGraph constructs a gated graph convolution network (GCN) framework based on an attention mechanism, capable of capturing both short-term and long-term patterns in dynamic graphs, and introduces negative sampling and edge loss functions to achieve anomaly edge detection. In addition, the TADDY method introduces a Transformer network architecture, simultaneously modeling the spatiotemporal features of dynamic graphs, effectively improving the performance of dynamic graph anomaly detection. Although dynamic graph anomaly detection has achieved good results in the above-mentioned fields, research in the power system domain is still in its early stages, with limited existing research achievements. Current research on dynamic graph-based anomaly detection in the power system domain mainly focuses on the detection of false data injection attacks (FDIAs). For example, the DynWatch-Local model uses a neighborhood knowledge-driven graph distance measurement method to dynamically estimate the measurement data distribution of the current power grid state based on historical data for anomaly detection. The TGNN method combines graph neural networks and gated recurrent networks (GRUs) to effectively detect and locate false data injection attacks in power systems. The GGNN method uses an attention mechanism to integrate power system operation data and topology information, extracting the spatiotemporal features of nodes to improve detection effectiveness. However, the above methods mainly focus on telemetering data, and the study of abnormal detection of power system topology connection anomalies, i.e., telesignaling data, has not been fully developed.

[0006] Due to the high dependence of power system measurement data on the dynamically changing power grid topology and load state, the context dependence is strong, and the accurate modeling of the conditional distribution of the measurement data faces great challenges. If the existing method ignores the background knowledge and physical law of the power grid, it is easy to cause high false positives, thereby reducing the reliability of the actual application. In addition, most of the existing methods are limited to a single scale in feature extraction, and it is difficult to efficiently capture complex spatio-temporal features of local and global scales in dynamic graphs, especially in complex and variable scenarios of abnormal events, the possibility of false negatives or false positives is high. SUMMARY

[0007] The purpose of the present application is to overcome the shortcomings and deficiencies of the prior art, and provide a power system abnormal remote signaling detection method based on a graph autoencoder model, which can effectively mine the potential structural patterns in power system data by fusing the graph autoencoder model and deep learning technology, comprehensively integrate dynamic spatio-temporal features and power grid physical knowledge, realize sensitive perception and accurate identification of remote signaling abnormalities, and provide an efficient, intelligent and physically interpretable new solution for remote signaling abnormal detection and structure repair in power systems.

[0008] To achieve the above purpose, the technical scheme provided by the present application is: a power system abnormal remote signaling detection method based on a graph autoencoder model, comprising the following steps:

[0009] S1, collecting measurement data of a power system and constructing it into graph data with node and edge relationships, simulating two types of remote signaling abnormalities in the graph, called edge abnormalities, i.e. false break abnormalities and false connection abnormalities, generating a labeled graph data set containing normal edges and abnormal edge samples through abnormal injection and labeling, and finally dividing it into a training set and a test set;

[0010] S2, inputting the graph data in the training set into the designed graph autoencoder model for training, the encoder part of the model realizes joint modeling of graph structure and time series information through residual enhanced graph attention mechanism and node-edge joint attention mechanism, maintains node representation stability and extracts key structural features, and combines a bidirectional long short-term memory network to capture the dynamic time series changes of the nodes, fuse multi-scale time features, and generate robust edge embedding representation; the above edge embedding representation is input into the graph decoder part of the model for edge feature reconstruction and abnormal score prediction, the graph decoder combines the physical prior of the power system to construct a physically constrained embedding space, realizes edge attribute reconstruction through inter-edge self-attention mechanism, and generates an abnormal score reflecting the degree of deviation of the structure or attribute; a joint loss function is used for optimization during training, including reconstruction loss, Kirchhoff current balance loss and classification loss for anomaly detection;

[0011] S3. During the training phase, the distribution of abnormal scores is statistically analyzed based on normal edge samples, and the threshold is optimized with the precision and recall of the test set as the optimization targets. During the testing phase, edge features are reconstructed and abnormal scores are scored on the input graph data. Abnormal edges are determined in combination with the threshold, and preliminary abnormal detection results are output.

[0012] S4. The detected abnormal edges are input into the graph convolution repair module. This module integrates graph topology, node features and edge attribute information. Through the neighborhood information propagation mechanism, it corrects the connection status and electrical parameters of the abnormal edges. The repaired edges can accurately restore the real topology of the power grid, effectively eliminate remote signaling errors, and improve the data quality of the power grid and the reliability of the power system operation.

[0013] Further, in step S1, measurement data of the power system is collected, and the collected measurement data is constructed into graph data with node and edge relationships. Two typical remote signaling anomalies, called edge anomalies, are simulated or identified in the graph data, including false disconnection anomalies and false connection anomalies. Anomaly injection, labeling, and normalization preprocessing are performed on the graph data to finally form a graph dataset with anomaly labels, that is, the graph dataset contains normal edge samples and abnormal edge samples. The graph dataset is divided into training set and test set. The features of the nodes include node voltage amplitude, generator active power, generator reactive power, load active power, and load reactive power. The features of the edges include active power PF at the starting side, reactive power QF at the starting side, active power PT at the ending side, reactive power QT at the ending side, and power system state estimation measurement data in the graph adjacency matrix. The graph data is represented as a time series graph set. T′ represents the total time, where the graph structure at each moment is defined as graph G. t ={V t E t}, where V t Let E represent the set of nodes in the power system at time t. t Let v represent the set of edges at time t; nodes represent buses, and edges represent transmission branch connections, reflecting the operating topology and state of the power system at time t; each node v∈V t It includes time-series characteristics such as node voltage magnitude, generator active power, generator reactive power, load active power, and load reactive power; each edge e∈E t Active power PF including the starting and ending points t PT t With reactive power QF t QT t The topological connectivity of the power grid is determined by the adjacency matrix. express, This represents the power grid topology at time t, specifically whether there is a connection between node i and node j. denotes that there is a connection between node i and node j at time t, denotes that there is no connection between node i and node j at time t, and time t indicates that the adjacency matrix is the structure description of graph G t , which embodies the dynamic changes of graph structure over time.

[0014] Further, in step S2, during the training process, the graph autoencoder model first models the structure and time sequence of the graph data through the encoder, adopts a residual enhanced graph attention encoder, combines a node-edge joint attention mechanism based on a graph attention network to extract key feature information in the graph structure, and maintains the stability of the original node representation, then introduces a bidirectional long short-term memory network to capture the dynamic change characteristics of the nodes in the time dimension, fuse different scales of time sequence information, and obtain more robust edge embedding representation. The residual enhanced graph attention encoder in the graph neural network layer is designed for power system graph structure data, which effectively captures abnormal information by fusing node-edge joint feature representation and historical state memory, which includes a node-edge joint attention mechanism based on a graph attention network, dynamic residual feature enhancement, cross-attention historical state reservation mechanism and time sequence information extraction module, which are introduced as follows:

[0015] ①Node-edge joint attention mechanism based on graph attention network:

[0016] To capture the relevance of power line abnormalities and the state of the two end nodes, an improved graph attention mechanism is designed, i.e., a node-edge joint attention mechanism based on a graph attention network, which explicitly fuses node attributes and topological relationships through the node-edge joint attention mechanism, as follows:

[0017] Given the feature vector x i ∈R d of the i-th node and the feature vector e ij ∈R e of the edge connecting node i and node j, where R is the real number field, d is the dimension of each node feature vector, and e is the feature dimension of each edge. First, feature alignment is performed through linear projection:

[0018]

[0019]

[0020] In the formula, W h is the weight matrix of node features; b h is the bias term of node features; W e is the edge feature weight matrix; and b e is the edge feature bias term. is the normalized feature of node i. The edge feature after nonlinear activation; LayerNorm normalizes the node feature to eliminate the dimensional difference; ReLU nonlinearly activates the edge feature;

[0021] The graph attention mechanism is used to interact features and calculate the attention weight of the edge:

[0022]

[0023] In the formula, represents the normalized attention weight of node j to node i in the lth attention head; a is a learnable attention vector, which measures the importance of node-edge joint features; l represents the lth attention head, and || represents the feature concatenation operation; represents the normalized feature of node i in the lth attention head; represents the normalized feature of node j in the lth attention head; represents the edge feature after nonlinear activation in the lth attention head;

[0024] Finally, the node representation h attn is obtained by multi-head aggregation:

[0025]

[0026] In the formula, L is the total number of attention heads, and N(i) is the neighbor set of node i;

[0027] ② Dynamic residual feature enhancement:

[0028] The topology of the power system changes dynamically. In order to avoid the degradation of deep network features, an adaptive residual connection mechanism is designed as follows:

[0029] Define the residual mapping function:

[0030]

[0031] In the formula, d in is the dimension of the input feature, i.e. the dimension of x i ; d out represents the dimension of the output feature, i.e. the dimension of the mapped W res x i ; W res is the weight matrix of residual mapping; F res is a dynamic residual mapping function, which can solve the key problems of dimension mismatch and gradient instability in deep learning model;

[0032] Through the gating mechanism, the attention features and residual features are dynamically fused:

[0033] g = σ(W g [hattn ||F res (x i )]);

[0034] h fusion =g⊙h attn +(1-g)⊙F res (x i );

[0035] In the formula, σ is a Sigmoid function, a smooth gate controls changes, and alleviates the problem of gradient disappearance in deep network; ⊙ represents element-wise multiplication; g is a gate vector, whose value is between [0, 1], the ratio of attention and residual information fusion obtained through the gate mechanism, used to control the dynamic fusion ratio of the two kinds of information, the greater g indicates that the current attention feature is more reliable, and the smaller g indicates that the history state needs to be relied on; h fusion is the fused feature vector, which combines attention information and residual information; W g is the weight matrix of the gate mechanism; this gate mechanism can adaptively adjust the fusion ratio of the original feature and the attention feature, and shows better adaptability in the scene of dynamic topology change of power system;

[0036] ③Cross-attention history state reservation mechanism:

[0037] In order to reserve the state evolution characteristics of the node itself, a cross-layer attention mechanism is introduced to adaptively select the key historical information and the current state fusion by taking the historical residual feature as a knowledge base; the output of the current layer is taken as Query, and the historical residual feature is taken as Key-Value:

[0038] Q=h fusion W q ;

[0039] K=F res (x i )W k ;

[0040] V=F res (x i )W v ;

[0041] The cross-layer attention weight β i is calculated:

[0042]

[0043] Through cross-layer attention and nonlinear change, the historical information and the current feature are fused to generate the final output h out :

[0044] h out =LayerNorm(ELU(β iV + h fusion ));

[0045] ELU activation function:

[0046]

[0047] ELU has a non-zero gradient in the negative interval, avoiding long-term non-updating of node features, and alleviating the Dead ReLU problem; the exponential term e x makes the negative value change smooth, suitable for scenarios where there are positive and negative values in power features; W q , W k , W v is the cross-attention query vector Q, key vector K, and value vector V projection weight; x is the input feature value; α is the scaling factor of the ELU function in the negative interval, controlling the amplitude of negative value mapping; the cross-attention history state reservation mechanism enables the node to adaptively reserve key historical state features when aggregating neighborhood information, enhancing the modeling ability of continuous evolution of nodes and edges;

[0048] ④ Time series information extraction module:

[0049] When detecting abnormal edges in power systems, it is found that the abnormal state of the edge is not only affected by the historical state of the node, but also triggered by the future context factors; therefore, a bidirectional long short-term memory network Bi-LSTM is introduced to enhance the modeling ability of time series context. Bi-LSTM considers both forward (history) and backward (future) information in the sequence, and can capture more comprehensive time-dependent features;

[0050] For each target edge, its state at multiple time steps is input into Bi-LSTM as an input sequence. Bi-LSTM is composed of two parallel LSTM layers, which perform information transmission along the positive and negative directions of time, respectively. The forward LSTM layer extracts historical context features, and the backward LSTM layer models the potential influence of future time. The hidden states output by the two layers are fused at each time step to form an edge time series representation containing bidirectional time series information.

[0051] The fused representation not only reflects the state of the edge at the current time, but also implicitly integrates the state change trend of the adjacent nodes within the time window before and after it, thereby providing more discriminative time series features for subsequent anomaly scoring.

[0052] Further, in step S2, the extracted edge embedding representation is input into a graph decoder for edge feature reconstruction and anomaly score prediction. The graph decoder based on physical prior is used to construct a physical constraint embedding space combined with the physical information of the power system, and the inter-edge self-attention mechanism is used to reconstruct the attribute features of the edges. The graph decoder based on physical prior includes a dynamic edge feature generation mechanism, an inter-edge self-attention mechanism and a physical prior fusion mechanism, which are specifically introduced as follows:

[0053] ① Dynamic edge feature generation mechanism:

[0054] The dynamic edge feature generation mechanism is selected based on the node dynamic generation of edge features, rather than directly relying on the original edge features. That is, the evolution of the node state is used to implicitly model the edge features to solve the problem of incomplete recording or noise interference of historical edge features. A dynamic edge feature generation module that combines the graph attention mechanism and the edge feature generator is proposed to dynamically generate edge features from node latent representations. The graph attention network dynamically aggregates neighborhood information to generate high-level semantic representations of nodes, and the edge feature generator dynamically generates edge features using the fusion features of source nodes and target nodes.

[0055] A multi-head GAT layer is used, and each head independently learns different attention weights. Finally, the multi-view features are integrated by splicing:

[0056]

[0057] In the formula, L is the total number of attention heads; l is the number of the lth attention head; sigma is a Sigmoid function, which is a smooth gate control change to alleviate the gradient disappearance problem of deep network; W l represents the learnable linear transformation weight matrix corresponding to the lth attention head, which maps the neighbor node features h j to the attention space for subsequent weighted summation; represents the normalized attention weight of node j to node i in the lth attention head; h i is the multi-head attention output feature of node i; || represents the feature splicing operation; N(i) is the neighbor set of node i; j∈N(i) represents that node j is a neighbor node of node i; h j represents the feature vector of node j, which is the neighbor node feature of node i;

[0058] The edge feature generator considers both the node high-level abstract features extracted by the GAT layer and the original latent features, thereby improving the fault tolerance of the model to sparse or noisy data:

[0059] e f =[h src ||h dst ||x src ||x dst];

[0060] where e f is the fused edge feature, h src is the h dst is the x src is the x dst concatenation result of the four features, which is used to represent the structure and attribute information comprehensively; h src is the node initial feature output by the GAT layer; h dst is the node final feature output by the GAT layer; x src is the original initial feature of the node; x dst is the original final feature of the node;

[0061] ② Inter-edge self-attention mechanism:

[0062] In order to effectively model the edge representation, an edge feature decoding module based on self-attention mechanism is designed. In the task of anomaly edge detection, the goal is to identify the edges in the power system that have potential faults or measurement errors. Since the flow measurement value of the edge may be missing or fake, the node state is used as a more robust feature to construct the edge representation.

[0063] The designed edge feature decoding module takes the state features of the two nodes connected by the edge as input, and models the interaction between the nodes through self-attention mechanism to obtain the edge representation containing context information.

[0064] Given the initial node feature e fs and the final node feature e ft connected by the edge, first, the two are concatenated to form the basic feature representation h base of the edge:

[0065] h base =[e fs ||e ft ];

[0066] In the edge feature decoding module, the self-attention mechanism is used to model the edge representation. The input of the self-attention mechanism is the basic feature representation h base , and the output of the self-attention mechanism is the edge representation h attn :

[0067] MultiheadAttn(Q,K,V)=Concat(head1,head2,…,head L )W o ;

[0068] Wherein, the attention mechanism adopts a multi-head attention mechanism, that is, the original input is decomposed into multiple heads, the attention weight corresponding to each head is calculated respectively, and then the attention weights obtained by each head are spliced to obtain the final attention weight; the Lth head W o is an output projection matrix, W L is the weight matrix of the Lth head, Attn is an attention function, Q, K and V are query, key and value vectors respectively;

[0069] Subsequently, h base and h attn are spliced and input into a multi-layer perception to obtain the final edge feature representation E pred :

[0070] E pred = MLP ([h base ||h attn ]) ;

[0071] ③ Physical prior fusion mechanism:

[0072] In order to enhance the model's perception of power physical information, a physical feature adapter is designed to map the original physical features to a hidden space representation compatible with the graph neural network, which is introduced as part of the node or edge input into the model, thereby providing prior information of the power grid operating state;

[0073] First, the physical features are used as input and mapped to a hidden space representation compatible with the graph neural network through a linear transformation. The physical feature adapter receives the original physical features of the nodes, including: voltage amplitude V m , node active power of generators and loads P, node reactive power of generators and loads Q, start point side active power PF, start point side reactive power QF, end point side active power PT and end point side reactive power QT, and generates hidden space physical features h phy through single-layer linear projection and nonlinear activation function Tanh, which is represented as:

[0074] h phy = Tanh (W adapter · [V m , P, Q, PT, PF, QT, QF] + b adapter ) ;

[0075] In the formula, W adapter and b adapter are learnable parameters, and the Tanh function constrains the feature range to [-1, 1] to avoid numerical divergence of physical quantities;

[0076] The hidden space physical features h phy are spliced with the edge feature representation Epred The fusion is performed to explicitly inject neighborhood knowledge constraints H:

[0077] H = MLP([E pred ,h phy ]);

[0078] In the execution process of the physical prior fusion mechanism, the prior information is implicitly injected through joint optimization of the feature space, the loss function imposes a penalty on predictions that violate physical laws, and backpropagation forces the physical feature adapter to learn feature representations that conform to domain knowledge. The graph features and physical hidden features are fused to generate predictions that conform to domain laws.

[0079] Further, in step S2, the edge feature reconstruction result output by the graph decoder and the anomaly score are jointly input to the joint loss function for optimization. The joint optimization objective includes three types of loss terms: reconstruction loss, which is used to improve the reconstruction accuracy of edge attributes; Kirchhoff current balance loss, which is used to ensure the physical consistency of the system state; and classification loss, which is used to optimize the anomaly detection performance according to the anomaly score;

[0080] The total loss function L' includes three terms:

[0081] L' = αL cls + βL recon + γL kcl ;

[0082] In the formula, α, β, and γ are hyperparameters that control the weights of the three constraints; L cls , L recon , and L kcl correspond to the classification loss, the reconstruction loss, and the Kirchhoff current balance loss, respectively. The specific descriptions of the classification loss, the reconstruction loss, and the Kirchhoff current balance loss are as follows:

[0083] ① Classification loss L cls uses Focal Loss to address the serious imbalance between abnormal edge samples and normal edge samples, and by strengthening the attention to abnormal edge samples, the model is prevented from being biased towards the majority class.

[0084] L cls = -Σ(1-p t ) γ log(p t );

[0085] In the formula, p t is the prediction probability of the model for the true class. If the sample is abnormal, it is called a positive sample, and p t is the probability that the model predicts as abnormal. If the sample is normal, it is called a negative sample, and p tis the probability that the model predicts as normal; γ is the focusing parameter to adjust the weight of easy and difficult samples, γ>0 reduces the weight of easy classification samples, forcing the model to focus on difficult classification samples, γ=0 degenerates into standard cross-entropy loss;

[0086] ② Reconstruction loss L recon Adopt physical constraint MSE to ensure that the power flow result predicted by the model conforms to the actual physical law of the power grid, prevent the model from fitting data only on statistical characteristics, and improve the generalization performance;

[0087]

[0088] Where, by minimizing the mean square error between the power flow predicted by the model and the real power flow e ij , the model output is forced to conform to the actual power system state;

[0089] ③ Kirchhoff current balance loss L kcl By explicitly introducing the node power conservation constraint, the physical consistency and credibility of the model prediction result are further strengthened, and the false detection caused by non-physical prediction is reduced. The goal of this loss is to convert the law of conservation of energy of the power system into a differentiable mathematical form and embed it into the training process of the machine learning model, so that the model not only learns the statistical law in the data, but also follows the underlying physical rules. In order to achieve this goal, the following core concepts need to be defined clearly: edge power flow, node power conservation and KCL constraint;

[0090] ① Edge power flow definition:

[0091] For edge connected nodes v i -v j , define the power flow feature:

[0092]

[0093]

[0094] In the formula, e ij represents the edge feature vector from node v i to node v j , which contains four components e ij [0], e ji [1], e ij [2], e ji [3], represents the active power output from node v i to v j , represents the active power input from node v j to v i , Indicates from node v i Flow to v j Active power output, Indicates from node v j Flow to v i The reactive power input; each edge e contains bidirectional active and reactive power flow information i→j and j→i;

[0095] ② Definition of nodal power conservation:

[0096] ΔP i =P i gen -P i load -P i in -P i out ;

[0097]

[0098] In the formula, ΔQ i For node v i The reactive power imbalance, ΔP i For node v i The active power imbalance, Indicates from node v i Flow to v j Active power output, Indicates from node v j Flow to v i Active power input, Indicates from node v i Flow to v j Active power output, Indicates from node v j Flow to v i reactive power input, P i gen For node v i The active power of the generator, P i load For node v i The active power of the load, For node v i The reactive power of the generator For node v i The reactive power of the load;

[0099] ③KCL constraint definition:

[0100] By forcing the model output to follow physical laws through node power balancing, anomaly false positives are reduced.

[0101]

[0102] In the formula, V is a node set.

[0103] Further, in step S3, after the training is completed, the abnormal score distribution is counted based on the normal edge samples in the training set, and a threshold is set as a subsequent determination basis. The threshold is selected to optimize the precision and recall on the test set. In the test phase, the graph data in the test set is input to reconstruct the edge features and calculate the abnormal scores. The set threshold is combined to determine the abnormal edges, and the preliminary abnormal edge detection result is output. The abnormal score is calculated from the reconstructed features of the graph decoder of the graph autoencoder model, which is used to measure the deviation of the edge in structure or attribute. The greater the reconstruction error, the more the edge deviates from the normal pattern in the graph structure, and thus it is considered as a potential abnormal edge.

[0104] Let the edge e=(s,v), and the abnormal score score(e) of the edge is the reconstruction error of the graph autoencoder model for the edge, that is:

[0105]

[0106] In the formula, is the edge feature reconstructed by the graph decoder part of the graph autoencoder model based on the physical prior; e f is the original edge feature; ||·||2 is the L2 norm; s is the starting node of the edge e, v is the terminal node of the edge e, and e=(s,v) is the edge connecting s and v; the abnormal score score(e) is used to measure the fitting ability of the graph autoencoder model for the edge structure, and the greater the error, the more likely the edge is an abnormal edge.

[0107] After obtaining the abnormal score of each edge, a threshold τ is set. When the abnormal score score(e) of a certain edge satisfies score(e)>τ, the edge is determined to be an abnormal edge. The selection of the threshold is optimized according to the performance indicators on the test set to balance the precision and recall.

[0108] Further, in step S4, for the abnormal edges determined in step S3, the graph repairing model uses a graph convolutional neural network to repair the abnormal edges, the graph convolutional neural network propagates node representations on the updated graph structure containing the abnormal edges, so that the node embeddings fuse semantic information of the surrounding context, and then the graph repairing model takes the repaired node representations as input, reconstructs the attributes of the abnormal edges by splicing the features of the nodes at both ends of the edges, and uses a multilayer perceptron to reconstruct the attributes of the abnormal edges; for the structural abnormal edges, if disconnected, the graph repairing model introduces candidate edges, and uses the updated node representations of the graph convolutional neural network to determine the rationality of the connections, and restores or removes the connections if necessary; finally, the graph repairing model outputs the repaired graph structure and edge features, realizes the joint repair of the telemetry data errors and structural abnormalities, and enhances the correctness of the measurement data of the power system.

[0109] Compared with the prior art, the present application has the following advantages and beneficial effects:

[0110] 1. The present application proposes a physical law constrained graph autoencoder model, which converts Kirchhoff's current law into a physical constraint loss, and combines it with the classification loss and reconstruction loss of the graph autoencoder to form a joint optimization objective, thereby achieving a synergistic improvement in abnormal detection accuracy and compliance with physical laws.

[0111] 2. The present application designs a residual-enhanced graph attention encoder, which constructs a node-edge joint attention mechanism based on a graph attention network, a dynamic residual feature enhancement, a cross-attention history state reservation mechanism, and a time series information extraction module coupled encoding structure, and effectively alleviates the deep feature degradation problem by fusing topology and features through the node-edge joint attention mechanism and combining the gating mechanism.

[0112] 3. The present application constructs a graph decoder based on physical priors, which introduces a physical adapter, a dynamic edge feature generation mechanism, and an inter-edge self-attention mechanism in the graph decoder, embeds physical feature information into the model hidden space, and improves the reconstruction ability and sensitivity to topological abnormalities.

[0113] 4. The present application fuses multi-modal and multi-scale feature representation capabilities, simultaneously uses graph structure features, node electrical quantities, and time series information for dynamic multi-scale modeling, and improves the generalization ability and robustness of the model in complex scenarios.

[0114] 5. The present application realizes end-to-end high-performance anomaly detection, significantly improves the accuracy of edge anomaly detection by introducing reconstruction error as an anomaly score and combining the physical constraint optimization objective, and significantly reduces the false positive rate.

[0115] 6、Based on IEEE118 standard power grid test set, the experiment proves that the graph autoencoder model proposed in the application shows excellent performance under various abnormal modes. In the experiment with 5% abnormal proportion, the model maintains a high accuracy of 98.7% and a precision of 96.5%, while achieving a recall rate of 92.3%, indicating that it can accurately distinguish normal and abnormal edges in a low abnormal proportion scenario while maintaining comprehensive coverage of abnormal situations. When the abnormal proportion increases to 10%, the performance of the model remains stable, with only a slight fluctuation in accuracy (98.5%) and precision (96.2%), and a recall rate of 91.0% verifies that it can maintain excellent detection coverage when the number of abnormalities increases. Even under the condition of 15% high abnormal proportion, the model still shows significant robustness, with an accuracy of 98.0%, a precision of 95.5%, and a recall rate of 89.8%, proving its reliable detection ability in a high noise environment. These experimental results fully verify that the model has accuracy, physical consistency and engineering adaptability in the abnormal edge remote signaling detection task of the power system, and its performance shows a gradient stability with the increase of abnormal proportion, providing high reliability support for the stable operation of the smart grid. BRIEF DESCRIPTION OF DRAWINGS

[0116] Figure 1 is a flowchart of the method of the application.

[0117] Figure 2 is an architectural diagram of the method of the application. DETAILED DESCRIPTION

[0118] The application will be described in further detail below with reference to the embodiments and the accompanying drawings, but the embodiments of the application are not limited thereto.

[0119] As shown in Figure 1 and Figure 2 , the embodiment discloses a power system abnormal remote signaling detection method based on a graph autoencoder model, and the specific circumstances are as follows:

[0120] S1, collect the measurement data of the power system, construct the collected measurement data into graph data with node and edge relationship, and simulate or identify two typical remote signaling abnormalities in the graph data, called edge abnormalities, including false break abnormality and false connection abnormality, perform abnormal injection, labeling and standardized preprocessing on the graph data, and finally form a graph data set with abnormal labels, that is, the graph data set contains normal edge samples and abnormal edge samples; divide the graph data set into a training set and a test set.

[0121] The characteristics of the nodes include node voltage amplitude, generator active power, generator reactive power, load active power and load reactive power, and the characteristics of the edges include start point side active power PF, start point side reactive power QF, terminal point side active power PT, terminal point side reactive power QT and power system state estimation measurement data in the graph adjacency matrix; the graph data is represented as a time series graph set T' represents total time, wherein the graph structure at each time is defined as a graph G t = {V t , E t}, wherein V t represents a node set of the power system at time t, E t represents an edge set at time t; the node represents a bus, and the edge represents a transmission branch connection relationship, reflecting the operation topology and state of the power system at time t; each node v∈V t contains node voltage amplitude, generator active power, generator reactive power, load active power and load reactive power time series characteristics; each edge e∈E t contains start point side and terminal point side active power PF t , PT t and reactive power QF t , QT t , and the topological connection state of the power grid is represented by an adjacency matrix , represents the power grid topology structure at time t, that is, whether there is a connection between node i and node j, represents that there is a connection line between node i and node j at time t, represents that there is no connection between node i and node j at time t, and time t indicates that the adjacency matrix is a structure description of the graph G t , embodying the dynamic change of the graph structure with time.

[0122] In this embodiment, the collected measurement data is specifically the NREL-118 test system dataset, which is based on the IEEE 118-bus standard test system reconstruction and integrates the generation capacity and load characteristics of the Western Electricity Coordinating Council (WECC) 2024 public case database. The dataset covers 8760 hours of synchronous time series data from January 1, 2024, 0 o'clock to December 31, 2024, 24 o'clock. The dataset contains 3 interconnected regions, 118 buses, 186 transmission lines, and 327 generating units, covering 9 types of power generation technologies (including thermal power, wind power, photovoltaic, etc.), and including hourly actual values (Real-Time, RT) and day-ahead forecast values (Day-Ahead, DA) of load, wind power output, and photovoltaic power output time series. The dataset supports deep coupling of synchronous time series data and power grid physical models, supporting dynamic topology anomaly detection and operating state inference.

[0123] For the NREL-118 test system dataset, we simulated three types of abnormal situations:

[0124] The first type is that the line should be connected but is actually disconnected: randomly select some edges from the original topology and remove them to create a line break anomaly. During the removal process, ensure that the overall network topology remains connected to avoid isolated nodes.

[0125] The second type is that the line should be disconnected but is actually connected: randomly insert new edges between node pairs that are not originally connected and randomly initialize their characteristics. The active power characteristics of the new edges follow a Gaussian distribution, and the reactive power characteristics are randomly sampled from the characteristics of existing normal edges.

[0126] The third type is line characteristic anomaly: randomly select some edges from the remaining lines that are normally operating and adjust their active power characteristics to ensure that the sum of the input / output power of the edges remains unchanged.

[0127] For this dataset, the overall anomaly ratio is set to 5%, 10%, and 15%, respectively, with line connection anomalies (should be connected but disconnected) accounting for 30%, line misconnection anomalies (should be disconnected but connected) accounting for 30%, and line characteristic anomalies accounting for 40%.

[0128] S2. Input the graph data in the training set into the designed graph autoencoder model for training. During training, the graph autoencoder model first performs joint structural and temporal modeling of the graph data through the encoder. It adopts a residual-enhanced graph attention encoder, combined with a node-edge joint attention mechanism based on a graph attention network to extract key feature information in the graph structure and maintain the stability of the original node representation. Subsequently, by introducing a bidirectional long short-term memory network, it captures the dynamic change features of nodes in the time dimension, and fuses temporal information at different scales to obtain a more robust edge embedding representation. The extracted edge embedding representation is input into the graph decoder for edge processing. Feature reconstruction and anomaly score prediction employ a physics-based graph decoder, which combines the physical information of the power system to construct a physical constraint embedding space. An edge attention mechanism is used to reconstruct the attribute features of edges, and anomaly scores are generated based on these reconstructed attribute features to measure the degree of deviation of edges from their structure or attributes. The edge feature reconstruction results and anomaly scores output by the graph decoder are input into a joint loss function for optimization. The joint optimization objective includes three types of loss terms: reconstruction loss, used to improve the reconstruction accuracy of edge attributes; Kirchhoff current balancing loss, used to ensure the physical consistency of the system state; and classification loss, used to optimize anomaly detection performance based on the anomaly scores.

[0129] The residual-enhanced graph attention encoder is a graph neural network layer designed for power system graph structure data. By fusing node-edge joint feature representation with historical state memory, it effectively captures abnormal information. It includes: a node-edge joint attention mechanism based on graph attention network, dynamic residual feature enhancement, a cross-attention historical state retention mechanism, and a time-series information extraction module, which are described in detail below:

[0130] ① Node-edge joint attention mechanism based on graph attention network:

[0131] To capture the correlation between power line anomalies and the states of nodes at both ends, this paper designs an improved graph attention mechanism, namely a node-edge joint attention mechanism based on graph attention networks. Traditional GNNs only model interactions between nodes, neglecting the direct influence of edge features. The node-edge joint attention mechanism explicitly fuses node attributes and topological relationships, as follows:

[0132] Given the feature vector x of the i-th node i ∈R d The eigenvector e of the edge connecting node i and node j ij ∈R e Where R is the real number field, d is the dimension of each node's feature vector, and e is the feature dimension of each edge; first, feature alignment is performed through linear projection:

[0133]

[0134]

[0135] where W h is the weight matrix of node features; b h is the bias term of node features; W e is the edge feature weight matrix; b e is the bias term of edge features. is the normalized feature of node i. is the nonlinear activated edge feature; LayerNorm normalizes the node feature to eliminate the dimension difference; ReLU nonlinearly activates the edge feature.

[0136] The graph attention mechanism is used to interact features and calculate the attention weight of the edge:

[0137]

[0138] where, represents the normalized attention weight of node j to node i in the lth attention head; a is a learnable attention vector, which is used to measure the importance of node-edge joint features; l represents the lth attention head, and || represents the feature concatenation operation. represents the normalized feature of node i in the lth attention head. represents the normalized feature of node j in the lth attention head. represents the nonlinear activated edge feature in the lth attention head.

[0139] Finally, the node representation h attn is obtained by multi-head aggregation.

[0140]

[0141] where L is the total number of attention heads, and N(i) is the neighbor set of node i.

[0142] ② Dynamic residual feature enhancement:

[0143] The topology of the power system changes dynamically, and the traditional fixed residual connection is difficult to adapt to the feature fusion requirements under different operating states. To avoid the degradation of deep network features, an adaptive residual connection mechanism is designed as follows:

[0144] Define the residual mapping function:

[0145]

[0146] where d in is the dimension of the input feature, i.e., the dimension of x i ; d outdimension of the output feature, i.e. W after mapping res x i dimension of W res is the weight matrix of residual mapping; F res is a dynamic residual mapping function, which can solve the key problems of dimension mismatch and gradient instability in deep learning models;

[0147] The attention feature and the residual feature are dynamically fused through the gating mechanism:

[0148] g = σ(W g [h attn || F res (x i )]);

[0149] h fusion = g o h attn + (1-g) o F res (x i );

[0150] In the formula, σ is a Sigmoid function, the smooth gate controls the change, and the gradient vanishing problem of deep network is alleviated; o represents element-wise multiplication; g is a gating vector, whose value is between [0, 1], the fusion ratio of attention and residual information obtained through the gating mechanism, which is used to control the dynamic fusion ratio of the two kinds of information, the greater g indicates that the current attention feature is more reliable, and the smaller g indicates that the history state needs to be relied on; h fusion is the fused feature vector, which combines the attention information and the residual information; W g is the weight matrix of the gating mechanism; compared with the fixed residual connection, this gating mechanism can adaptively adjust the fusion ratio of the original feature and the attention feature, and shows better adaptability in the scene of dynamic topology change of power system;

[0151] ③Cross-attention history state reservation mechanism:

[0152] In order to reserve the state evolution characteristics of the node itself, a cross-layer attention mechanism is introduced, which adaptively selects the key historical information and the current state fusion by taking the historical residual feature as a knowledge base; the output of the current layer is taken as Query, and the historical residual feature is taken as Key-Value:

[0153] Q = h fusion W q ;

[0154] K = F res (x i )W k ;

[0155] V = F res (x i )W v ;

[0156] Compute cross-layer attention weights β i :

[0157]

[0158] Fuse historical information with current features through cross-layer attention and nonlinear transformation to generate final output h out :

[0159] h out = LayerNorm(ELU(β i V + h fusion ));

[0160] ELU activation function:

[0161]

[0162] where ELU has non-zero gradient in negative interval to avoid node features from being updated for a long time and alleviate the Dead ReLU problem; the exponential term e x makes the negative value change smooth, which is suitable for scenarios with positive and negative values in power features (such as bidirectional power flow); W q , W k , and W v are projection weights of the query vector Q, the key vector K, and the value vector V of cross-attention; x is the input feature value; a is the scaling coefficient of the ELU function in the negative interval, which controls the amplitude of the negative value mapping; the cross-attention historical state reservation mechanism enables the node to adaptively reserve key historical state features when aggregating neighborhood information, thereby enhancing the modeling capability for continuous evolution of nodes and edges;

[0163] ④ Temporal information extraction module:

[0164] When detecting abnormal edges in power systems, it is found that the abnormal state of an edge is not only affected by the node state at the historical time, but also triggered by the future context factors. Therefore, a bidirectional long short-term memory network (Bi-LSTM) is introduced to enhance the modeling capability for time series context. Bi-LSTM considers both forward (i.e., historical) and backward (i.e., future) information in the sequence, and can capture more comprehensive time-dependent features.

[0165] For each target edge, its state at multiple time steps is input into Bi-LSTM as an input sequence. Bi-LSTM is composed of two parallel LSTM layers, which perform information transmission along the positive and negative directions of time, respectively. The forward LSTM layer extracts historical context features, while the backward LSTM layer models the potential influence of future time. The hidden states output by the two layers are fused at each time step to form an edge temporal representation containing bidirectional time series information.

[0166] The fused representation can not only reflect the edge's own state at the current time, but also implicitly integrate the state change trend related to the adjacent nodes within the time window before and after the current time, thereby providing more discriminative time sequence feature support for subsequent anomaly scoring.

[0167] The physical prior-based graph decoder includes a dynamic edge feature generation mechanism, an inter-edge self-attention mechanism, and a physical prior fusion mechanism, which are specifically introduced as follows:

[0168] ① Dynamic edge feature generation mechanism:

[0169] In the abnormal edge detection task of the power system, the edge feature usually changes dynamically over time. If the original feature corresponding to the edge is directly used, the problem of incomplete recording of historical edge features or interference by noise may be encountered. Therefore, we choose to generate edge features based on nodes rather than directly relying on original edge features, that is, we implicitly model the edge features through the evolution of node states to solve the problem of incomplete recording of historical edge features or interference by noise, and propose a dynamic edge feature generation module that fuses a graph attention mechanism and an edge feature generator to dynamically generate edge features from node latent representations. The graph attention network generates a high-level semantic representation of the node by dynamically aggregating neighborhood information, while the edge feature generator dynamically generates the features of the edge using the fused features of the source node and the target node.

[0170] A multi-head GAT layer is used, and each head independently learns different attention weights, and finally the multi-perspective features are integrated by splicing:

[0171]

[0172] In the formula, L is the total number of attention heads; l is the number of the lth attention head; σ is a Sigmoid function, a smooth gate control change, which alleviates the problem of gradient disappearance in deep networks; W l represents the learnable linear transformation weight matrix corresponding to the lth attention head, which maps the neighbor node features h j to the attention space for subsequent weighted summation. represents the normalized attention weight of node j to node i in the lth attention head; h i is the multi-head attention output feature of node i; || represents the feature splicing operation; N(i) is the neighbor set of node i; j∈N(i) represents that node j is a neighbor node of node i; h j represents the feature vector of node j, which is the neighbor node feature of node i.

[0173] The edge feature generator considers both the node high-level abstract features extracted by the GAT layer and the original latent features, thereby improving the fault tolerance of the model to sparse or noisy data:

[0174] e f =[h src ||h dst ||x src ||x dst ];

[0175] In the formula, e f To fuse edge features, h is src ||h dst ||x src ||x dst The concatenation result of the four features is used to comprehensively represent the structural and attribute information of the product; h src h represents the initial node features output by the GAT layer. dst The node termination feature output by the GAT layer; x src The original initial features of the node; x dst This refers to the original termination feature of the node;

[0176] ② Marginal self-attention mechanism:

[0177] To effectively model edge representation, an edge feature decoding module based on a self-attention mechanism is designed. In the abnormal edge detection task, the goal is to identify edges in the power system that have potential faults or measurement errors. Since the power flow measurement values ​​of the edges may be missing or falsified, the edge representation is constructed starting from the more robust node states.

[0178] Specifically, the edge feature decoding module takes the state features of the nodes at both ends of the edge as input, and explicitly models the interaction relationship between the nodes through a self-attention mechanism to obtain the edge representation containing contextual information.

[0179] Given the feature e of the starting node connected by the edge fs With the terminating node feature e ft First, the two are concatenated to form the basic feature representation h of the edge. base :

[0180] h base =[e fs ||e ft ];

[0181] In the edge feature decoding module, a self-attention mechanism is used to model the edge representation. The input of the self-attention mechanism is the basic feature representation h. base The output of the self-attention mechanism is an edge representation h. attn :

[0182] MultiheadAttn(Q,K,V)=Concat(head1,head2,...,headL )W o ;

[0183] wherein the attention mechanism adopts a multi-head attention mechanism, i.e., the original input is decomposed into multiple heads, the attention weight corresponding to each head is calculated respectively, and then the attention weights obtained by each head are spliced to obtain the final attention weight; the Lth head W o is an output projection matrix, W L is a weight matrix of the Lth head, Attn is an attention function, Q, K and V are query, key and value vectors respectively;

[0184] Subsequently, h base is spliced with h attn , and input into a multi-layer perception to obtain the final edge feature representation E pred :

[0185] E pred = MLP([h base ||h attn ]);

[0186] ③ Physical prior fusion mechanism:

[0187] In order to enhance the model's perception ability of power physical information, a physical feature adapter is designed to map the original physical features to a hidden space representation compatible with the graph neural network, which is introduced as part of the node or edge input into the model, thereby providing prior information of the power grid operating state;

[0188] Firstly, the physical features are used as input, and a linear transformation is used to map the physical features to a hidden space representation compatible with the graph neural network. The physical feature adapter receives the original physical features of the node, including: voltage amplitude V m , node active power P (including generator and load), node reactive power Q (including generator and load), branch flow characteristics (start point side active power PF, start point side reactive power QF, end point side active power PT and end point side reactive power QT), hidden space physical features h phy are generated through single-layer linear projection and nonlinear activation function (Tanh), which are represented as:

[0189] h phy = Tanh(W adapter ·[V m , P, Q, PT, PF, QT, QF] + b adapter );

[0190] In the formula, W adapter and b adapterAs learnable parameters, the Tanh function constrains the feature range to [-1, 1], avoiding the numerical divergence of physical quantities;

[0191] The physical feature adapter injects the physical prior knowledge into the latent space by fusing the edge feature representation E phy with the physical latent feature h pred :

[0192] H=MLP([E pred ,h phy ]);

[0193] During the execution of the physical prior fusion mechanism, the prior information is implicitly injected through the joint optimization of the feature space. The loss function imposes a penalty on predictions that violate physical laws, and backpropagation forces the physical feature adapter to learn feature representations that conform to domain knowledge. By fusing graph features and physical latent features, the model generates predictions that conform to domain laws.

[0194] To solve the multi-constraint problem of abnormal edge detection in power systems, a joint optimization objective is designed, which includes three constraints: reconstruction loss for improving the reconstruction accuracy of edge attributes; Kirchhoff's current balance loss (KCL) for ensuring the physical consistency of system state; classification loss for optimizing abnormal detection performance according to abnormal scores;

[0195] The total loss function L' contains three terms:

[0196] L′=αL cls +βL recon +γL kcl ;

[0197] where α, β, and γ are hyperparameters that control the weights of the three constraints; L cls , L recon , and L kcl correspond to the classification loss, reconstruction loss, and Kirchhoff's current balance loss, respectively. The specific descriptions of the classification loss, reconstruction loss, and Kirchhoff's current balance loss are as follows:

[0198] ① The classification loss L cls uses Focal Loss to address the severe imbalance between abnormal edge samples and normal edge samples. By emphasizing the attention to abnormal edge samples, the model avoids biasing towards the majority class.

[0199] L cls =-∑1-p t ) γ log(p t );

[0200] where p t is the predicted probability of the true class. If the sample is abnormal, it is called a positive sample, and pt is the probability that the model predicts the sample as abnormal; if the sample is normal, called a negative sample, then p t is the probability that the model predicts the sample as normal; γ is a focusing parameter to adjust the weights of easy and difficult samples, γ > 0 reduces the weight of easy-to-classify samples, forcing the model to focus on difficult-to-classify samples, and γ = 0 degenerates into the standard cross-entropy loss;

[0201] ② Reconstruction loss L recon Adopt physical constraint MSE to ensure that the power flow result predicted by the model conforms to the actual physical law of the power grid, prevent the model from fitting data only on statistical characteristics, and improve the generalization performance;

[0202]

[0203] where the mean square error between the power flow predicted by the model and the real power flow e ij is minimized, forcing the model output to conform to the actual power system state;

[0204] ③ Kirchhoff's current balance loss L kcl By explicitly introducing the node power conservation constraint, the physical consistency and credibility of the model prediction result are further strengthened, and the false detection caused by non-physical prediction is reduced. The goal of this loss is to convert the law of conservation of energy of the power system into a differentiable mathematical form and embed it into the training process of the machine learning model, so that the model not only learns the statistical law in the data, but also follows the underlying physical rules. In order to achieve this goal, the following core concepts need to be defined: edge power flow, node power conservation and KCL constraint;

[0205] ① Edge power flow definition:

[0206] For edge-connected nodes v i -v j , define the power flow feature:

[0207]

[0208]

[0209] In the formula, e ij represents the edge feature vector from node v i to node v j , which contains four components e ij [0], e ji [1], e ij [2], e ji [3], represents the active power output from node v i to v j , represents the reactive power output from node vj flow to v i , the active power input, flow from node v i flow to v j , the active power output, flow from node v j flow to v i , the reactive power input; each edge e contains bidirectional active and reactive power flow information i→j and j→i;

[0210] ②Node power conservation definition:

[0211] ΔP i = P i gen - P i load - P i in - P i out ;

[0212]

[0213] where ΔQ i is the reactive power imbalance at node v i , ΔP i is the active power imbalance at node v i , flow from node v i flow to v j , the active power output, flow from node v j flow to v i , the active power input, flow from node v i flow to v j , the active power output, flow from node v j flow to v i , the reactive power input, P i gen is the generator active power at node v i , P i load is the load active power at node v i , is the generator reactive power at node v i , is the load reactive power at node v i ;

[0214] ③KCL constraint definition:

[0215] The output of the node power balance forced model complies with physical laws, thereby reducing abnormal misjudgments:

[0216]

[0217] In the formula, V is a node set.

[0218] S3, after training, the abnormal score distribution of the normal edge samples in the training set is counted, and a threshold is set as a subsequent determination basis, and the threshold is selected to optimize the precision and recall on the test set; in the test stage, the graph data in the test set is input to reconstruct the edge features and score the abnormality, and the set threshold is combined to determine the abnormality of the edge, and the preliminary abnormal edge detection result is output.

[0219] The calculation of the abnormal score is derived from the reconstructed features of the edge by the graph decoder of the graph autoencoder model, and the greater the reconstruction error, the more the edge does not conform to the normal mode in the graph structure, and thus is regarded as a potential abnormal edge;

[0220] Let the edge e=(s,v), then the abnormal score score(e) of the edge is the reconstruction error of the edge by the graph autoencoder model, that is:

[0221]

[0222] In the formula, is the edge feature reconstructed by the graph decoder part of the graph autoencoder model based on physical prior; e f is the original edge feature; ||·||2 is the L2 norm; s is the starting node of the edge e, v is the terminal node of the edge e, e=(s,v) is the edge connecting s and v; the abnormal score score(e) is used to measure the fitting ability of the graph autoencoder model to the edge structure, and the greater the error, the more likely the edge is an abnormal edge;

[0223] After obtaining the abnormal score of each edge, a threshold τ is set, when the abnormal score score(e) of a certain edge satisfies score(e)>τ, it is determined that the edge is an abnormal edge; wherein the selection of the threshold is optimized according to the performance indicators on the test set to balance the precision and recall.

[0224] S4, input the output abnormal edge detection result into the graph repair module, the graph repair module is designed based on graph convolutional neural network (GCN), receives graph structure, node embedding and reconstructed edge feature as input, in the case of abnormal edge connection or attribute deviation, the context information of the node representation is fused through the graph convolution propagation mechanism, and the attribute or connection state of the abnormal edge is reconstructed based on the representation of the nodes at both ends of the edge, and finally the repaired edge structure and edge feature are output, the destroyed remote signaling state in the power grid is restored, and the integrity of the graph structure and the operation reliability of the power system are improved.

[0225] For the edges determined to be abnormal, the graph repair model uses a graph convolutional neural network to repair, which propagates node representations on the updated graph structure containing abnormal edges, so that the node embedding fuses the semantic information of the surrounding context. Subsequently, the graph repair model takes the repaired node representation as input, reconstructs the attributes of the abnormal edge by concatenating the features of the nodes at both ends of the edge, and uses a multilayer perceptron; for the structural abnormal edge, if disconnected, the graph repair model introduces candidate edges and uses the updated node representation of the graph convolutional neural network to determine the rationality of the connection, and restores or removes the connection if necessary; finally, the graph repair model outputs the repaired graph structure and edge features, realizes the joint repair of telemetry data errors and structural abnormalities, and enhances the correctness of the power system measurement data.

[0226] The experimental results of the above-mentioned power system abnormal remote signaling detection method based on the graph autoencoder model of the embodiment will be described in detail below:

[0227] According to the final detection results, the performance is evaluated from the accuracy, precision and recall. In order to evaluate the performance of the present application, an ablation experiment is performed. By comparing with the model without the time series information extraction module, the model without the node-edge joint attention mechanism based on the graph attention network, the model without the dynamic residual feature enhancement and cross attention mechanism, and the model without the graph decoder based on the physical prior, the influence of each module on the accuracy, precision and recall of the model is analyzed. For the three abnormal proportions 5%, 10% and 15% of the NREL-118 test system data set, the corresponding comparison test is performed. The comparison results of 5% are shown in Table 1.

[0228] Table 1

[0229]

[0230] The above table results show that in the experiment of 5% abnormal proportion, the present application still maintains high accuracy 98.7% and precision 96.5%, showing that the present application can effectively distinguish normal and abnormal edges under low abnormal proportion, and the recall rate 92.3% shows that it can cover all abnormal situations well. In contrast, the model after removing the module shows a large performance decline, especially in the recall rate. After removing the time series information extraction module, the recall rate decreases significantly (85.6), indicating that the time series feature is crucial for power system anomaly detection. After removing the physical prior graph decoder, the precision and recall rate further decrease, indicating that the introduction of physical constraints is crucial for the accuracy of anomaly detection. The comparison results of 10% are shown in Table 2.

[0231] Table 2

[0232]

[0233] The above table results show that in the experiment of 10% abnormal proportion, the application still shows significant advantages, the stability of accuracy rate 98.5% and precision rate 96.2% is verified, and the recall rate 91.0% is higher, which shows that the application can still comprehensively detect the anomaly under more abnormal conditions. In contrast, removing the time sequence information extraction module causes the recall rate to drop to 82.3%, proving the importance of time sequence features in the dynamic power grid environment. The decrease in precision rate and recall rate caused by removing the node-edge joint attention mechanism and the physical prior decoder highlights the necessity of these mechanisms in the model. After removing the dynamic residual feature enhancement module, the model accuracy decreases slightly, but the recall rate is relatively stable, indicating that it has less impact on the overall performance.

[0234] The 15% comparison results are shown in Table 3.

[0235] Table 3

[0236]

[0237] The above table results show that in the experiment of 15% abnormal proportion, the performance of the application is gradually affected by more anomalies, but the application still shows strong robustness and superiority, with accuracy rate 98.0% and precision rate 95.5%, and recall rate 89.8%, indicating that under high abnormal proportion, the application can effectively identify abnormal edges with low false positive rate. After removing the time sequence information extraction module, the recall rate drops sharply to 78.9%, indicating that time sequence information is more important for anomaly detection under high abnormal proportion. After removing the node-edge joint attention mechanism and the physical constraint decoder, the precision rate and the recall rate both decrease significantly, highlighting the indispensability of these modules in high abnormal detection. Removing the dynamic residual feature enhancement module has less impact on the recall rate, but the precision rate decreases.

[0238] The above embodiments are the preferred embodiments of the application, but the embodiments of the application are not limited by the above embodiments, and any changes, modifications, substitutions, combinations, simplifications made without departing from the spirit and principles of the application shall be equivalent replacement methods and shall be within the scope of protection of the application.

Claims

1. A power system abnormal telecommunication detection method based on a graph autoencoder model, characterized in that, The method comprises the following steps: S1, collecting measurement data of the power system and constructing the measurement data into graph data with node and edge relationships, simulating two types of remote signaling abnormalities in the graph, referred to as edge abnormalities, namely false break abnormalities and false connection abnormalities, generating a labeled graph data set containing normal edges and abnormal edge samples through abnormal injection and labeling, and finally dividing the labeled graph data set into a training set and a test set; S2, inputting the graph data in the training set into a designed graph autoencoder model for training, the encoder part of the model realizing joint modeling of graph structure and time sequence information through residual enhanced graph attention mechanism and node-edge joint attention mechanism, maintaining node representation stability and extracting key structural features, and combining a bidirectional long short-term memory network to capture dynamic time sequence changes of nodes, fuse multi-scale time features, and generate robust edge embedding representation; the above edge embedding representation is input into the graph decoder part of the model for edge feature reconstruction and abnormal score prediction, the graph decoder combines power system physical priors to construct a physically constrained embedding space, realizes edge attribute reconstruction through inter-edge self-attention mechanism, and generates abnormal scores reflecting the deviation degree of structure or attribute; a joint loss function is used for optimization during training, including reconstruction loss, Kirchhoff current balance loss and classification loss for anomaly detection; S3, in the training stage, the abnormal score distribution is statistically analyzed based on normal edge samples and the threshold is optimized, and the precision and recall rates of the test set are used as the optimization target; in the test stage, edge feature reconstruction and abnormal scoring are performed on the input graph data, and the abnormal edges are determined based on the threshold, and the preliminary anomaly detection result is output; S4, the detected abnormal edges are input into a graph convolution repair module, which integrates graph topology, node features and edge attribute information, corrects the connection state and electrical parameters of the abnormal edges through a neighborhood information propagation mechanism, and the repaired edges can accurately restore the true topology relationship of the power grid, effectively eliminate remote signaling errors, and improve the operation reliability of the power grid.

2. The power system abnormal telecommunication detection method based on a graph autoencoder model according to claim 1, wherein, In step S1, measurement data of the power system is collected, and the collected measurement data is constructed into graph data with node and edge relationships. Two typical remote signaling anomalies, called edge anomalies, are simulated or identified in the graph data, including false disconnection anomalies and false connection anomalies. Anomaly injection, labeling, and normalization preprocessing are performed on the graph data to finally form a graph dataset with anomaly labels, that is, the graph dataset contains normal edge samples and abnormal edge samples. The graph dataset is divided into training set and test set. The features of the nodes include node voltage amplitude, generator active power, generator reactive power, load active power, and load reactive power. The features of the edges include active power PF at the starting side, reactive power QF at the starting side, active power PT at the ending side, reactive power QT at the ending side, and power system state estimation measurement data in the graph adjacency matrix. The graph data is represented as a time series graph set. T ′ Let G represent the total time, where the graph structure at each moment is defined as graph G. t ={V t E t }, where V t Let E represent the set of nodes in the power system at time t. t Let v represent the set of edges at time t; nodes represent buses, and edges represent transmission branch connections, reflecting the operating topology and state of the power system at time t; each node v∈V t It includes time-series characteristics such as node voltage magnitude, generator active power, generator reactive power, load active power, and load reactive power; each edge e∈E t Active power PF including the starting and ending points t PT t With reactive power QF t QT t The topological connectivity of the power grid is determined by the adjacency matrix. express, This represents the power grid topology at time t, specifically whether there is a connection between node i and node j. This indicates that there is a connecting line between node i and node j at time t. This indicates that there is no connection between node i and node j at time t, and time t indicates that the adjacency matrix is ​​graph G. t The structural description reflects the dynamic changes of the graph structure over time. 3.The power system abnormal telecommunication detection method based on the graph auto-encoder model according to claim 1, wherein, In step S2, during the training process, the graph autoencoder model first models the structure and time sequence of the graph data through the encoder, adopts a residual enhanced graph attention encoder, combines a node-edge joint attention mechanism based on a graph attention network to extract key feature information in the graph structure, and maintains the stability of the original node representation, then captures dynamic change features of the nodes in the time dimension by introducing a bidirectional long short-term memory network, fuses time sequence information of different scales, and obtains more robust edge embedding representation, wherein the residual enhanced graph attention encoder is a graph neural network layer designed for power system graph structure data, which effectively captures abnormal information by fusing node-edge joint feature representation and historical state memory, and includes a node-edge joint attention mechanism based on a graph attention network, a dynamic residual feature enhancement mechanism, a cross-attention historical state reservation mechanism and a time sequence information extraction module, which are specifically introduced as follows: ①Node-edge joint attention mechanism based on graph attention network: To capture the relevance between power line anomalies and the state of nodes at both ends, an improved graph attention mechanism, i.e., a node-edge joint attention mechanism based on a graph attention network, is designed. Through the node-edge joint attention mechanism, the node attributes and topological relationships are explicitly fused, as follows: the feature vector of the i-th node x i ∈ R d and the feature vector of the edge connecting node i and node j e ij ∈ R e where R is the field of real numbers, d is the dimension of each node feature vector, and e is the feature dimension of each edge; first, feature alignment is performed by linear projection: In the formula, W h is a weight matrix of node features; b h is a bias term of node features; W e is the edge feature weight matrix; b e is the bias term for edge features; is the normalized feature for node i; is the edge feature after nonlinear activation; LayerNorm normalizes node features to eliminate dimension differences; ReLU activates edge features nonlinearly; The graph attention mechanism is used for feature interaction to calculate the attention weight of the edge: In the formula, represents the normalized attention weight of node j to node i in the lth attention head; a is a learnable attention vector, which is used to measure the importance of node-edge joint features; l represents the lth attention head, and || represents a feature splicing operation; represents the normalized feature of node i in the lth attention head; represents the normalized feature of node j in the lth attention head; represents the edge feature after nonlinear activation in the lth attention head; Finally, the node representation h is obtained by multi-head aggregation attn : where L is the total number of attention heads, and N(i) is the neighbor set of node i. ②Dynamic residual feature enhancement: The topology of the power system changes dynamically. To avoid feature degradation in deep networks, an adaptive residual connection mechanism is designed, as follows: Define the residual mapping function: In the formula, d in is the dimension of input features, i.e. x i is the dimension of input features; d out represents the dimension of output features, i.e. the dimension of W res x i after mapping; W res is the weight matrix of residual mapping; F res is a dynamic residual mapping function, which can solve the key problems of dimension mismatch and gradient instability in deep learning models; The attention features and residual features are dynamically fused through a gating mechanism: g = σ(W g [h attn ||F res (x i )]) h fusion = g O h attn + (1 - g) O F res (x i ); In the formula, σ is a Sigmoid function, a smooth gate control change, and a deep network gradient vanishing problem is alleviated; ⊙ represents element-by-element multiplication; g is a gate vector, the value of which is between [0, 1], the attention and residual information fusion ratio obtained through the gate mechanism is used to control the dynamic fusion ratio of the two kinds of information, and the greater g indicates that the current attention feature is more reliable, and the smaller g indicates that the history state needs to be relied on; h fusion is a fused feature vector, which combines attention information and residual information; W g is a weight matrix of the gate mechanism; the gate mechanism can adaptively adjust the fusion ratio of the original feature and the attention feature, and shows better adaptability in the dynamic topology change scene of the power system. ③Cross-attention historical state reservation mechanism: To preserve the state evolution characteristics of the node itself, a cross-layer attention mechanism is introduced. The historical residual features are used as a knowledge base to adaptively select key historical information and fuse it with the current state. The output of the current layer is used as Query, and the historical residual features are used as Key-Value: Q = h fusion W q ; K = F res (x i )W k ; V = F res (x i )W v ; Computing cross-layer attention weights β i : By cross-layer attention and nonlinear transformation, historical information is fused with current features to generate the final output h out : h out = LayerNorm(ELU(β i V + h fusion )); ELU activation function: In the formula, ELU has a non-zero gradient in the negative interval, avoids long-time non-updating of node features, and alleviates the Dead ReLU problem; the exponential term e x Smooths the negative value change and is suitable for a scene in which there are positive and negative values in power features; W q , W k , W v is the cross-attention query vector Q, key vector K, and value vector V projection weight; x is the input feature value; alpha is the scaling coefficient of the ELU function in the negative interval, controls the amplitude of the negative value mapping; the cross-attention history state reservation mechanism enables the node to adaptively reserve key historical state features when aggregating neighborhood information, and enhances the modeling capability for continuous evolution of nodes and edges; ④Temporal information extraction module: When detecting abnormal edges in the power system, it is found that the abnormal state of an edge is not only affected by the node state at the historical time, but also triggered by the future context. Therefore, a bidirectional long short-term memory network (Bi-LSTM) is introduced to enhance the modeling capability of the time series context. Bi-LSTM considers both the forward (historical) and backward (future) information in the sequence, and can capture more comprehensive time-dependent features. For each target edge, its state at multiple time steps is input into Bi-LSTM as an input sequence. Bi-LSTM consists of two parallel LSTM layers that perform information transmission in the forward and backward directions, respectively. The forward LSTM layer extracts historical context features, and the backward LSTM layer models the potential impact of future time steps. The hidden states output by the two layers are fused at each time step to form an edge temporal representation that contains bidirectional time series information. The fused representation not only reflects the edge's own state at the current time, but also implicitly integrates the state change trend of its adjacent nodes within the time window before and after it, thereby providing more discriminative time series features to support subsequent anomaly scoring.

4. The power system abnormal telecommunication detection method based on a graph autoencoder model according to claim 1, wherein, In step S2, the extracted edge embedding representation is input into the graph decoder for edge feature reconstruction and anomaly score prediction. A graph decoder based on physical priors is used to combine the physical information of the power system to construct a physical constraint embedding space and reconstruct the attribute features of the edge using an inter-edge self-attention mechanism. The graph decoder based on physical priors includes a dynamic edge feature generation mechanism, an inter-edge self-attention mechanism, and a physical prior fusion mechanism, which are described as follows: ①Dynamic edge feature generation mechanism: The edge feature is dynamically generated based on the node state, rather than directly relying on the original edge feature. The problem of incomplete or noisy historical edge features is solved by implicitly modeling the edge feature through the evolution of the node state. A dynamic edge feature generation module is proposed, which combines graph attention mechanism and edge feature generator to dynamically generate edge features from node latent representations. The graph attention network generates high-level semantic representations of nodes by dynamically aggregating neighborhood information, while the edge feature generator dynamically generates edge features using the fusion features of source nodes and target nodes. Multiple GAT layers are used, each learning different attention weights independently. The multi-view features are finally integrated by concatenation: where L is the total number of attention heads; l is the number of the l-th attention head; σ is a Sigmoid function, which smoothes the gate control change and alleviates the gradient vanishing problem of deep networks; W l represents the learnable linear transformation weight matrix corresponding to the l-th attention head, which acts to map the neighbor node features h j to the attention space for subsequent weighted summation; represents the normalized attention weight of node j to node i in the l-th attention head; h i is the multi-head attention output feature of node i; || represents the feature concatenation operation; N(i) is the neighbor set of node i; j∈N(i) represents that node j is a neighbor node of node i; h j represents the feature vector of node j, which is the neighbor node feature of node i; The edge feature generator considers both the high-level abstract features extracted by the GAT layer and the original latent features, thereby improving the model's fault tolerance to sparse or noisy data: e f = [h src ||h dst ||x src ||x dst ]; In the formula, e f is the fusion edge feature, h src ||h dst ||x src ||x dst The spliced results of the four features are used to comprehensively represent the structure and attribute information; h src is the node starting feature output by the GAT layer; h dst is the node ending feature output by the GAT layer; x src is the original node starting feature; x dst is the original node ending feature; ②Inter-edge self-attention mechanism: To effectively model the edge representation, a self-attention-based edge feature decoding module is designed. In the anomaly edge detection task, the goal is to identify edges in the power system that have potential faults or measurement errors. Since the power flow measurements of edges may be missing or fake, the edge representation is constructed based on the more robust node state. The designed edge feature decoding module takes the state features of the two nodes at the ends of the edge as input and explicitly models the interaction between nodes through self-attention mechanism to obtain an edge representation that contains context information. The starting node feature e connected by the given edge fs The ending node feature e ft First, the two are spliced to form the basic feature representation h of the edge base : h base = [e fs ||e ft ]; In the edge feature decoding module, the edge representation is modeled using a self-attention mechanism, the input of the self-attention mechanism is the base feature representation h base , and the output of the self-attention mechanism is the edge representation h attn : MultiheadAttn(Q, K, V) = Concat(head1, head2,..., head L )W o ; Wherein, the attention mechanism adopts a multi-head attention mechanism, that is, the original input is decomposed into multiple heads, the attention weight corresponding to each head is calculated respectively, and then the attention weights obtained by each head are spliced to obtain the final attention weight; the Lth head W o is an output projection matrix, W L is a weight matrix of the Lth head, Attn is an attention function, Q, K and V are query, key and value vectors respectively; Subsequently, h base is concatenated with h attn is concatenated, and a multi-layer perceptron is input to obtain the final edge feature representation E pred : E pred = MLP([h base || h attn ]); ③Physical prior fusion mechanism: To enhance the model's perception of physical information, a physical feature adapter is designed to map the original physical features to a hidden space representation compatible with graph neural networks. This representation is introduced as part of the node or edge input to the model, providing prior information about the state of the power grid. First, using physical features as input, and mapping the physical features to a hidden space representation compatible with the graph neural network through a linear transformation, the physical feature adapter receives the original physical features of the nodes, including: voltage amplitude V m , the node active power P of the generator and the load, the node reactive power Q of the generator and the load, the starting point side active power PF, the starting point side reactive power QF, the terminal point side active power PT and the terminal point side reactive power QT, the hidden space physical features h phy generated by single-layer linear projection and nonlinear activation function Tanh are represented as: h phy = Tanh(W adapter · [V m , P, Q, PT, PF, QT, QF] + b adapter ); where W adapter and b adapter are learnable parameters, and the Tanh function constrains the feature range to [-1, 1] to avoid numerical divergence of physical quantities. The physical feature adapter injects the latent space physical feature h phy with the edge feature representation E pred to fuse to explicitly inject neighborhood knowledge constraints H: H = MLP([E pred ,h phy ]) ; During the execution of the physical prior fusion mechanism, the prior information is implicitly injected through joint optimization of the feature space. The loss function penalizes predictions that violate physical laws, and backpropagation forces the physical feature adapter to learn feature representations that conform to domain knowledge. The fusion of graph features and physical hidden features generates predictions that conform to domain rules.

5. The power system abnormal telecommunication detection method based on a graph autoencoder model according to claim 1, wherein, In step S2, the edge feature reconstruction result output by the graph decoder is input into the joint loss function together with the anomaly score for optimization. The joint optimization objective includes three types of loss terms: reconstruction loss to improve the reconstruction accuracy of edge attributes; Kirchhoff current balance loss to ensure the physical consistency of system state; and classification loss to optimize the anomaly detection performance based on the anomaly score. Total loss function L ′ Comprises three terms: L' = aL cls + bL recon + gL kcl ; where a, b and g are hyperparameters that control the weights of the three constraints, respectively; L cls , L recon and L kcl correspond to the classification loss, reconstruction loss and Kirchhoff current balance loss, respectively; the specific introductions of the classification loss, reconstruction loss and Kirchhoff current balance loss are as follows: ①Classification loss L cls Focal Loss is used to deal with the serious imbalance between abnormal edge samples and normal edge samples. By strengthening the attention to abnormal edge samples, the model is prevented from being biased towards the majority class. L cls = -∑(1 - p t )γlog(p t ); where p t is the predicted probability of the model for the real class, and if the sample is abnormal, called positive sample, then p t is the probability predicted by the model as abnormal. If the sample is normal, called negative sample, then p t is the probability of the model predicting normal; γ is the focusing parameter to adjust the weight of difficult samples, γ > 0 reduces the weight of easy classification samples, forcing the model to focus on difficult classification samples, γ = 0 degenerates into standard cross-entropy loss; ② Reconstruction loss L recon The physical constraint MSE is adopted to ensure that the power flow results predicted by the model comply with the actual physical laws of the power grid, prevent the model from only fitting the data in statistical characteristics, and improve the generalization performance. where the model is forced to output the actual power system state by minimizing the mean squared error between the model predicted power flow and the true power flow e ij ​ ③Kirchhoff's current balance loss L kcl By explicitly introducing the node power conservation constraint, the physical consistency and reliability of the model prediction results are further strengthened, and the false detection caused by non-physical prediction is reduced. The goal of this loss is to convert the law of conservation of energy of the power system into a differentiable mathematical form and embed it into the training process of the machine learning model, so that the model not only learns the statistical rules in the data, but also follows the underlying physical rules. In order to achieve this goal, the following core concepts need to be defined: edge power flow, node power conservation and KCL constraint; ①Edge power flow definition: For edge-connected node v i -v j , define the power flow feature: In the formula, e ij Indicates from node v i Pointing to node v j The edge feature vector contains 4 components e ij [0]、e ji [1], e ij [2]、e ji [3], Indicates from node v i Flow to v j Active power output, Indicates from node v j Flow to v i Active power input, Indicates from node v i Flow to v j Active power output, Indicates from node v j Flow to v i The reactive power input; each edge e contains bidirectional active and reactive power flow information i→j and j→i; ②Node power conservation definition: ΔP i = P i gen - P i load - P i in - P i out ; In the formula, ΔQ i For node v i The reactive power imbalance, ΔP i For node v i The active power imbalance, Indicates from node v i Flow to v j Active power output, Indicates from node v j Flow to v i Active power input, Indicates from node v i Flow to v j Active power output, Indicates from node v j Flow to v i reactive power input, P i gen For node v i The active power of the generator, P i load For node v i The active power of the load, For node v i The reactive power of the generator For node v i The reactive power of the load; ③KCL constraint definition: The node power balance forces the model output to follow physical laws, reducing false positives: where V is the set of nodes.

6. The power system abnormal telecommunication detection method based on a graph autoencoder model according to claim 1, wherein, In step S3, after training, the anomaly score distribution is calculated based on the normal edge samples in the training set, and a threshold is set as the basis for subsequent judgment. The threshold is selected to optimize the precision and recall on the test set. In the test phase, the graph data in the test set is inputted for edge feature reconstruction and anomaly scoring, and the edges are determined to be abnormal in combination with the set threshold, and the preliminary abnormal edge detection result is outputted; wherein the abnormal score is calculated from the reconstructed feature of the edge by the graph decoder of the graph autoencoder model, and is used to measure the deviation degree of the edge in structure or attribute. The greater the reconstruction error is, the more the edge does not conform to the normal mode in the graph structure, and thus is regarded as a potential abnormal edge; Let the edge e=(s,v), and the abnormal score score(e) of the edge is the reconstruction error of the edge by the graph decoder of the graph autoencoder model, i.e. wherein, is the edge feature reconstructed by the physics-prior based graph decoder part in the graph autoencoder model; e f is the original edge feature; ||·||2is the L2 norm; s is the start node of edge e, v is the end node of edge e, e = (s, v) is the edge connecting s and v; the anomaly score score(e) is used to measure the fitting ability of the graph autoencoder model to the edge structure, and the larger the error is, the more likely the edge is an abnormal edge; After obtaining the abnormal score of each edge, a threshold τ is set. When the abnormal score score(e) of an edge satisfies score(e)>τ, the edge is determined to be abnormal. The selection of the threshold is optimized according to the performance index on the test set to balance the precision and recall.

7. The power system abnormal telecommunication detection method based on a graph autoencoder model according to claim 1, wherein, In step S4, for the edges determined to be abnormal, the graph repair model uses the graph convolutional neural network to repair. The graph convolutional neural network propagates the node representation on the updated graph structure containing the abnormal edge, so that the node embedding fuses the semantic information of the surrounding context. Subsequently, the graph repair model takes the repaired node representation as input, concatenates the features of the two nodes of the edge, and uses the multilayer perceptron to reconstruct the attribute of the abnormal edge. For the structural abnormal edge, if the connection is disconnected, the graph repair model introduces candidate edges, and uses the updated node representation of the graph convolutional neural network to determine the rationality of the connection, and restores or removes the connection if necessary. Finally, the graph repair model outputs the repaired graph structure and edge feature, realizes the joint repair of telemetry data errors and structural abnormalities, and enhances the correctness of the power system measurement data.

Images