Research On Detection Method Of Power System False Data Injection Attack Based On Machine Learning

On January 17,2019,State Grid Corporation of China raised the construction and operation of the “Ubiquitous Power Internet of Things” to the strategic position of the company in the report of the two conferences.As the degree of intelligence of China's power system is further enhanced,the degree of damage caused by cyber attacks may be Exceeded normal expectations.The power system already has the typical characteristics of the Icyber-Physical System(CPS),and the occurrence of information security incidents may cause serious consequences such as large-scale blackouts.As a new type of power system network attack,false data injection attacks(FDIAs)can successfully bypass the bad data detection mechanism,offset the power measurement data,mislead the control center operation under extremely hidden conditions,and seriously threaten the stability of the power system.run.The traditional detection method is difficult to detect such attacks.In order to ensure the security of the smart grid,this paper analyzes the false data injection attack mechanism,utilizes the advantages of machine learning in dealing with the two classification problems,and supervises the learning method to inject false data into the attack.The detection work is summarized in two steps: model training and classification decision.The research on the detection method of false data injection attack of power grid is carried out from three aspects: data sample construction,feature extraction and detection model construction.This paper builds an attack detection model based on machine learning.Then the power measurement data sample set containing positive and negative samples is the basis of model training and detection experiments.First,consider the two conditions that the attacker grasps the complete grid topology information and the non-complete grid topology information.The attack mechanism of the false data injection attack is discussed in depth.Then,the network topology of the IEEE-14-bus and IEEE-118-bus standard node systems is constructed and the power measurement data is generated.It is difficult for the attacker to master the complete power in practice.System parameters and network topology.This paper constructs FDIAs attack vectors under the condition of non-complete network topology information.Through simulation experiments,normal and negative data samples for attack detection are generated,which is the implementation of further attack detection model.Provide the basis.Power measurement data has high dimensionality and strong noise characteristics,and it is difficult to directly apply to model training and detection experiments.Data reduction methods for measurement data alone cannot guarantee the targeting of attack detection,using Isolated Forests(iForest)and Locally Linear Embedding.(LLE)Advantages in anomaly detection and data dimensionality reduction,innovatively combining abnormal score extraction and data dimensionality reduction,and proposing a feature extraction method for iForest-LLE power measurement data specifically for FDIAs detection.It not only ensures the targeting of attack detection in the data processing stage,but also takes into account the data quality of feature extraction.Finally,the effectiveness and superiority of the proposed feature extraction method are verified by experiments.The primary goal of FDIAs attack detection is to ensure accuracy.It is difficult to detect well-designed FDIAs based on the theory of power system alone.This paper designs an attack detection model based on Gradient Boosting Decision Tree(GBDT),which uses basic decision tree algorithm and gradient.The combination of the promotion framework enables a single decision tree classification model to continuously improve the accuracy in serial training,so as to construct a high-precision attack detection model;the problem of the GBDT model hyperparameter selection has a large impact on the results and is difficult to determine.Based on the original fruit fly optimization algorithm,a three-dimensional adaptive chaotic fruit fly algorithm was proposed.The automatic optimization strategy of hyperparameters in the training process was designed.The basic GBDT detection model was improved and recorded as V3ACFOA-GBDT.Simulation experiments show that This method has advantages in the computational efficiency and detection accuracy of attack detection.