Attack Detection Of Synchrophasor Measurement Data In Power System Considering Spatio-temporal Correlations

The modern power system has been more intelligent and automatic with the rapid development of communication technology,and has already possessed the typical characteristics of CPS(Cyber-physical System).However,the deep coupling of the cyber side and the physical side of the power system has gradually increased the risk of cyber security of the cyber side which will spread among the physical side.Specifically,PDMAs(PMU Data Manipulation Attacks)will cause the control center to not correctly obtain the real-time operating status of the power system,which seriously threatens the safety and stability of the system.Traditional attack detection methods can no longer meet the current defense needs of the power system.This paper sets out to carry out research on the detection methods of PDMA based on deep learning technologies.The specific work and results are as follows:Firstly,a method for detecting attacks in synchrophasor measurements of power system based on LSTM(Long Short-Term Memory)is proposed.According to the characteristics of PMU data changing with time,this method uses LSTM models to extract temporal features of PMU time series data.It is a typical data-mining method.After learning the features of normal data,the prediction errors of the model is used to judge whether the features of real-time data conforms to the normal ones,that is,whether there is a data manipulation attack.Therefore,the model is trained only by positive samples,and solves the problem of imbalance between positive and negative PMU data samples.Simulation results show that the method can effectively detect PDMAs and has a high accuracy rate.Secondly,a DCAE(Denoising Convolutional Autoencoder)-based power system synchrophasor measurement data attack detection method is proposed.This method points to the entire monitoring network of the power system and learns the spatial relationship between multiple PMU measurement data based on the CAE(Convolutional Autoencoder)model.This model learns the characteristics of normal samples,reconstructs the real-time data input,detects the data whose actual value deviates from the reconstructed output value as tampered data,and realizes PDMAs detection.In combination with the fact that PMU data is often exposed to noise in engineering application,a DCAE model is further proposed to denoise.After comparative analysis of multiple models,the DCAE model not only has a better detection performance for measurement data of multiple PMUs,but also has strong robustness in noisy environments.Finally,an attack detection method based on LSTM-DSCAE for synchrophasor measurement data of power system considering spatiotemporal correlations is proposed.This method draws lessons from the pre-training process of SAE(Stacked Autoencoder).It integrates and improves the above two methods,and uses Dropout techniques to prevent overfitting.Because the measurement data of the power system has complex spatiotemporal correlations,this method can not only extract deeper and more abstract hidden features,but also have higher learning efficiency and better detection capability.Simulation results show that the method has higher accuracy than some supervised learning methods,and has a more comprehensive feature extraction capability compared with LSTM and DCAE models.Moreover,this method lows the possibility of the misjudgment of power system fault data to a certain extent.