Applications of integral reinforcement learning control in electrical machines and power converter systems

• Main Author: Yu, Yang
• Other Authors: Su Rong
• Format: Thesis-Doctor of Philosophy
• Language: English
• Published: Nanyang Technological University 2021
• Subjects: Engineering::Electrical and electronic engineering::Control and instrumentation::Control engineering; Engineering::Electrical and electronic engineering::Power electronics
• Online Access: https://hdl.handle.net/10356/145783
• Institution: Nanyang Technological University

With the rapid development of power electronic devices, the ideas of more-electric aircraft (MEA) have become practical with many of the traditional hydraulic and pneumatic systems being replaced with equivalent electrical systems. MEA also motivates the development of segmented machines to betteutilize the available space inside the gas turbine. The stator of a segment machine is non-asymmetrically distributed so that existing pipes and cables of a gas turbine engine can be accommodated in the con ned space of this gas turbine engine. The model of segmented machine is normally derived numerically based on finite element analysis (FEA). Considering the data-driven nature of segmented machine, a promising control method called integral reinforcement learning (IRL) has been gaining attention to deal with the control problems associated with the segmented machine. Living organisms learn by acting on their environment, observing the resulting reward stimulus. IRL controller uses an actor/critic structure and adapts its control actions based on interactions with the data driven model. The idea of IRL based control combines the appealing features of optimal control theory and adaptive control theory. The practical application chapter of this thesis describes how IRL has been successful experimentally implemented into H infinity control of 2-kW electrical machine, which shows the superiority of an IRL controller over other classical controllers with the distinct advantage of learning an optimal control law without prior knowledge of the system model. However, relevant technical improvements are needed to make IRL more computationally efficient. To this end, few theoretical contributions are made in theoretical chapters.