Research On PMU Data Attack And Detection Method In Wide Area Measurement System

Wide area measurement system(WAMS),based on synchronous phasor measurement technology,aims to detect,analyze,and control the dynamic process of system for real-time monitoring of the power system.The system has the characteristics of high accuracy,high speed,and synchronous measurement in different places,and it has good adaptability for the wide area power grid.The system comprises phasor measurement units(PMUs),phasor data concentrators(PDCs),a data communication network,and a system control center.PMU utilizes a high-precision global positioning system(GPS)satellite synchronized clock to achieve synchronized measurements of grid bus voltage and line current phasor quantities.The synchronized measurements support power system dynamic analysis,including oscillation detection,load model and transmission line parameter identification,transmission and distribution state estimation,etc.However,there are various security issues in the process of collection and transmission of field PMU data.Malicious attackers can easily exploit them to launch different network attacks,thus triggering oscillations or power interruptions in the primary power system and seriously endangering the regular operation of the power system.To this end,this thesis takes the characteristics and vulnerabilities of PMU measurement as a starting point to profoundly analyze the principle of typical PMU data attacks and study the intelligent PMU data attack detection model based on neural networks.On this basis,the thesis introduces adversarial analysis and proposes a security and robustness protection method for intelligent detection models.Thus,the reliability of the measurement data obtained from the system control center is ensured in terms of both data and model security.The research work and main results of this thesis are as follows:(1)The industry standards of synchronous measurement adopted by WAMS lack a perfect network security mechanism so that attackers can easily use it to launch PMU data attacks,which can corrupt or manipulate synchronous measurements and affect the dynamic monitoring and analysis of the power system.To address this problem,the thesis analyzes the potential security risks of PMU data sampling and transmission in WAMS,introduces system observable conditions and measurement phasors critical conditions to give PMU sites vulnerability conditions under different measurement cases,and also constructs a vulnerability index in a cartesian coordinate system based on the measurement phasors reported by PMU sites to quantify the vulnerability degree of the sites.(2)The vulnerability of the WAMS system exposes the PMU measurements to cyber attacks,and attackers manipulate the system state estimation to trigger error control by introducing malicious data in the system that match the power characteristics and hardly increase the measurement residuals.To address this problem,the thesis constructs a WAMS measurement model based on PMU measurements,gives the unified basic principle of PMU data attack,with two typical attacks as examples,and gives a sparse attack matrix construction method with limited system network information.(3)A vector neural network-based false phase detection method is proposed to address the potential PMU data attacks in WAMS.The method focuses on the information contained in the phase volume of PMU measurements,"phase encodes"the amplitude and phase angle inside the phase volume,uses a dynamic routing mechanism to maintain the encoding relationship when passing information between layers,and searches for the hierarchical relationship between node encapsulation instances and the data vector as a whole,so as to detect the attacked data from a large number of normal measurement vectors.(4)For the intelligent detection model of PMU data attack in WAMS,an adversarial attack of sample feature saliency weakening is proposed from the attacker’s perspective which interferes with the detection process of the model by introducing an adversarial perturbation in the original PMU data attack sample,thus changing the detection result with high confidence and improving the success rate of the original PMU data attack.The attack analyzes the classification and detection process of PMU data by the intelligent detection model and deduces the contribution of each part of PMU data to the detection result in reverse,so as to introduce the perturbation in a targeted manner and achieve the attack transfer of the adversarial PMU data attack sample from the white-box model to the target black-box model.(5)For the adversarial PMU data attacks,an adversarial attack defense method based on causal theory is proposed to enable intelligent detection models to identify potential causal relationships in the observed data,so as to obtain causal encoding from sample data to labels and improve the robustness of the models.The method introduces causal structure learning to explore causal relationships in sample data,and uses potential causal relationships between samples,outputs and real labels to mitigate the impact of adversarial attacks,thus maintaining the accuracy of the detection model under different adversarial attack scenarios.