Research On The Detection And Mitigation Method Of False Data Injection Attacks In Power Grid Based On Deep Learning

The intelligence of the modern power grid has made it a cyber-physical system,but the massive application of information and communication technology and supervisory control and data acquisition system in the power grid has brought new dangers to the power grid.State estimation is one of indispensable modules of the power grid.It estimates the state of the power grid through measurements and is equipped with a bad data detection mechanism to eliminate abnormity.It is very this module that has a loophole,and an attacker can use the communication network to tamper with the measurements cleverly,and then can control the estimation without being detected.This type of network attack is called false data injection attack.In response to that,this paper aims to propose an anomaly detection method based on deep learning from the perspective of time series anomaly detection to plug the loophole and maintain the security of the smart grid.Power system DC state estimation is introduced at the beginning,followed by the bad data detection mechanism based on residuals,then the loopholes is analyzed,based on that,the principle of false data injection attack is derived.It is because false data injection attack does not change the estimation residual that it can pass the detection.To be trained better,it was decided to use the principle of constrained targeted attack to generate false data,and two scenarios that ordinary targeted attacks and playback attack were considered.Later,the concepts of attack duration and interval is added.The "sliding window" method was used to intercept sequences for model training and testing.This dissertation then proposes a detection method based on deep learning.The false data injection attack detection problem is regarded as a time series anomaly detection problem.Based on the idea of "to predict first,to judge later",a deep neural network was used to build a detector,which has a separate structure of "predictor+discriminator".The predictor is composed of a convolutional neural network and a recurrent neural network.The former is responsible for spatial feature extraction,and the latter is responsible for temporal feature extraction.The main body of the discriminator is a multi-layer perceptron.This paper designs three feature extraction method for discriminator:concatenation,convolution,and squared error.Besides,based on the prediction function of the detector,a data repair mechanism is proposed to ensure the accuracy of estimation and the accuracy of prediction.Finally,through simulation on the IEEE-39 bus system,the feasibility and effectiveness of the proposed mechanism are verified.It is found that only ordinary targeted attacks can be used for training to achieve a good result that various evaluation indexes are about 99%.