Research On Graph-Based False Data Detection In Power System

With the development of sensor technology and communication technology,the power system is deeply integrated with the above-mentioned technologies to form a power information physical system.However,in the process of data measurement and data transmission,the possibility of power system being attacked is getting higher and higher.Grid attacks are mainly aimed at disrupting the availability,integrity,and privacy of information.Currently,the main attack on the power grid is the false data injection attack,which constructs attack vectors based on the state estimation principle and tampers with the data during data transmission,destroying the integrity of the power grid operation data.False data injection attacks can evade the bad data detection mechanism of the control center state estimation system and have strong concealment.If the control center uses abnormal data for state estimation and applies it to subsequent tasks,it will have an impact on the safe and stable operation of the power system.This article constructs a false data detection model based on graphs,which is suitable for detecting abnormal data in dynamic and static power grid diagrams.Applying deep learning theory and graph neural network to construct anomaly data detection model based on graph supervised learning and graph unsupervised learning.Compared with data-driven detection methods,graph based detection models can effectively improve detection performance in dynamic topology situations.Firstly,the power system state estimation,bad data detection mechanisms,and false data injection attacks were studied;Analyzed the principles of deep learning,gated graph neural networks,attention mechanisms,and graph convolutional neural networks,which are key methods for spatial feature extraction of power grid information;Analyzed the internal process of node feature aggregation and the generation process of datasets.Secondly,aiming at the scene of sufficient abnormal data labels in the power grid,a false data injection attack detection model based on graph based supervised learning is proposed.The gated graph neural network is used for spatial feature extraction of grid node features and topological information,and an attention mechanism is added to the node feature aggregation process to assign reasonable aggregation weights to different neighboring nodes,which makes the accuracy of node representation greatly improved.Graph based models can accurately perceive changes and improve detection performance in dynamic topology situations.The effectiveness of this model was verified through numerical simulation.Finally,a false data injection attack detection model based on graph unsupervised learning is proposed for the lack of abnormal data labels in the power grid.Based on the graph self encoder,the graph convolutional neural network and attention mechanism are used to construct the encoder and obtain the hidden layer feature vector,respectively construct the structure decoder and attribute decoder to reconstruct the corresponding data,and calculate the reconstruction error to guide the update of network parameters.The graph unsupervised learning detection model is also applicable to dynamic topology scenarios.The feasibility of unsupervised model for anomaly data detection was verified through numerical simulation.However,the detection performance of the graph unsupervised learning model is lower than that of the graph with supervised learning detection model.If the abnormal data labels of the power grid can be obtained,it is more suitable to use the graph with supervised learning detection model,which is suitable for different detection scenarios.