LEARNING IN NONLINEAR SYSTEM CONTROL AND OPTIMIZATION

Abstract

Learning can significantly enhance the ability of engineer to deal with practical nonlinear problems since learning is a process whereby a system alters its action to perform a task more effectively due to the increase in knowledge related to the task. The purpose of the thesis is to apply the self-learning techniques to solve nonlinear system control and optimization problems. There are a number of issues to address in designing advanced control systems. One problem receiving increasing attention is the design of a controller when the output of the plant needs to be tightly controlled all the time, along a prespecified trajectory, such as robot manipulator for trajectory tracking. An approach to solve this problem is iterative learning control. Iterative learning control is a technique to improve the performance of systems or processes that operate repetitively over a fixed time interval. Conventional iterative learning control uses the error information obtained in the previous trial to Learn the new control input for the system. However, this is an open-loop control with respect to the current cycle operation. Hence its ability of disturbance rejection is limited. In this thesis, current cycle feedback is firstly adopted in the iterative learning algorithm, which can provide satisfactory rejection to external distributions or perturbations since the on-line information is soon used to adjust the control action and the feedback gain of this scheme can be theoretically chosen high enough to speed up the convergence speed of error. However, in practical implementation, the choice of feedback gain is sensitive to the sampling delay and the actual gain is bounded. Hence, a hybrid type iterative learning control algorithm, which incorporates D (derivative) type feedback from the previous trial as well as P (proportional) type feedback from the current trial, is proposed to provide not only satisfactory system robustness but also high convergence speed. Besides, the system using this type of learning algorithm is inherently stable regardless the feedback gain. In practice, this defines a broad class of systems to which the technique can be applied, even when the systems are characterized by imprecise knowledge of their dynamic behavior, as well as possible changes in operation environment. Another important issue for nonlinear system analysis and design is the optimization problem, which is usually multimodal and discontinuous. In this thesis, genetic-based learning algorithms are developed to search the optimal structure of fuzzy-neural network structure and the optimization solution of constrained power system. By using the self-learning techniques, nonlinear system iterative learning control algorithm and genetic-based optimization are well developed. Simulation results and case studies are also carried out and they show superior performance of the proposed algorithms.