Research On False Data Injection Attack Detection Method Of Smart Grid Based On Convolutional Neural Network

The development of smart grids has brought certain advantages to the development of power systems,but the integration of communication networks and the Internet of Things with the grid also brings network threats.Various network attacks can potentially harm the security,reliability,and economy of the grid,and can even lead to outage incidents and cascading failures.False Data Injection Attack(FDIA),as a common cyber attack,can introduce arbitrary false data into certain state variables to mislead the control center to make wrong judgments.Since FDIA is highly concealed and difficult to detect,the damage caused cannot be ignored.Therefore,it is urgent to design a high-precision FDIA detection method to lay the foundation for FDIA defense.This paper focuses on the detection method of FDIA.in smart grid,which mainly includes the following two parts.Convolutional Neural Network(CNN)is combined with Bidirectional Long And Short Term Memory(Bi-LSTM)and Bidirectional Gated Recurrent Unit(Bi-GRU),respectively,to design two hybrid structures of CNN-Bi-LSTM and CNN-Bi-GRU for false data detection,in which CNN adopts layer-by-layer structure to extract data features,Bi-LSTM and Bi-GRU capture past and future information respectively,and finally output the hidden state of each time step to fully utilize effective information of time series data to improve the final extraction effect of features.The dataset is validated using measurements from the Phasor Measurement Unit(PMU)and simulated using different FDIA detection methods,mainly including five deep learning methods,CNN-Bi-LSTM,CNN-Bi-GRU,CNN,LSTM and CNN-LSTM,three machine learning algorithms of Logistic Regression(LR),Support Vector Machines(SVM)and K Nearest Neighbor(KNN).The results show that for different levels of FDIA attack strength,both CNN-Bi-LSTM and CNN-Bi-GRU models consistently maintain high detection accuracy and are insensitive to the effect of noise,and still have stable FDIA detection performance when the measurement data contain severe noise.However,CNN-Bi-GRU has higher detection accuracy and better robustness.A parallel CNN detection method based on image data is proposed.Combining three conversion methods,namely Gramian Angular Field(GAF),Markov Transition Field(MTF),and Recurrence Plot(RP),the relevant features of power system time series measurement data are encoded into twodimensional images to obtain potential data features.A parallel convolutional neural networks(PCNN)model is designed to solve the problem of poor classification accuracy of a single CNN near the classification threshold and improve the feature extraction ability of high-dimensional data.The image data is used as the input of PCNN.After feature extraction,the output results determine whether there is FDIA in the power grid.The measurement data set from PMU is used to compare with the onedimensional CNN model and the single-channel two-dimensional CNN model respectively.The results show that the two-dimensional PCNN model proposed in this paper always maintains high detection accuracy for multiple data sets with different degrees of attack intensity.