Date of Award
Master of Science
David Yu, Guangwu Xu
Nowadays, multilevel power inverters have become a hot research topic which are being widely used in smart grids. They are also driving devices for conveyors, compressors, motors, and can enable uninterruptible power supply for critical loads such as database centers or telecommunications base stations. In the future, smart grids will play an important role to achieve higher efficiency, smarter control and better performance. Such an ambitious goal can only be achieved by inverters with higher voltage and power levels.
The smart grids are the typical cyber-physical systems that is composed of physical processes and computation units combined by sensors, actuators, and communication devices. The smart grids are apt to errors and vicious attacks on their physical construction leading to considerable damage, such as false data injection (FDI), denial of service (DOS). The vicious data injection can effectively bypass the detection of system and cause serious effects on the grid.
In recent years, some advanced control approaches have been proposed to perform inverter current control. Among them, model predictive control (MPC) is a promising one that makes use of explicit system models to predict its future response and optimize system performance. It has unique advantages that can accurately forecast the future response of the system and have fast response.
However, the effectiveness and the accuracy of the conventional MPC rely on whether the system model is accurate. Uncertainty and false data injection in the system model sometimes lead to unresponsive or even unstable control systems. Conventional MPC is hard to keep the system stable when the uncertainty and malicious attack happen. In existing studies, although various attacks have been investigated, the undetectable false data injection aiming at the inverter system was rarely studied.
In the thesis, the model of the cascaded H-bridge inverter is established and conventional MPC to achieve load current control is applied. It shows great performance to achieve load current control and has fast dynamic control. Then considering various attack signals such as step attack signals, pulse attack signals to the sensors in the system, the conventional MPC loses the ability to achieve a stable and effective current control.
According to simulation results, Kalman Filter model is built which can filter some Gaussian noises from the sensors in the system. Then from the perspective of attacker, a special FDI attack is designed that can effectively bypass the Kalman Filter. For the system that targeted by the FDI and DOS attack, a new controller is designed based on the K-Nearest Neighbor (KNN) algorithm and MPC strategy which can achieve the load current control with high output quality. Finally, the new control method based on KNN and MPC is compared with conventional MPC. The simulation results are analyzed and conclusion have been made. A modified MPC combined with KNN algorithm proposed in this thesis can detect bad data that can enter the system without triggering alarms. The case studies show the modified MPC based on KNN algorithm can achieve current control accurately when the system is injected by various attack signals showing better performance of current control with low total harmonic distortion (THD).
Rao, Rao, "Model Predictive Control for Mitigating Sensor Attacks on Multilevel Inverters" (2019). Theses and Dissertations. 2237.