Control and Estimation Oriented Model Order Reduction for Linear and Nonlinear Systems

Model Order Reduction;Naval Architecture and Marine Engineering;Engineering;Naval Architecture & Marine Engineering

Choroszucha, RichardEustice, Ryan M ;

University of Michigan

关键词: Model Order Reduction;    Naval Architecture and Marine Engineering;    Engineering;    Naval Architecture & Marine Engineering;   

Others  :  https://deepblue.lib.umich.edu/bitstream/handle/2027.42/140839/riboch_1.pdf?sequence=1&isAllowed=y

瑞士|英语

来源: The Illinois Digital Environment for Access to Learning and Scholarship

【 摘 要 】

Optimization based controls are advantageous in meeting stringent performance requirements and accommodating constraints.Although computers are becoming more powerful, solving optimization problems in real-time remains an obstacle because of associated computational complexity. Research efforts to address real-time optimization with limited computational power have intensified over the last decade, and one direction that has shown some success is model order reduction.This dissertation contains a collection of results relating to open- and closed-loop reduction techniques for large scale unconstrained linear descriptor systems, constrained linear systems, and nonlinear systems.For unconstrained linear descriptor systems, this dissertation develops novel gramian and Riccati solution approximation techniques.The gramian approximation is used for an open-loop reduction technique following that of balanced truncation proposed by (Moore, 1981) for ordinary linear systems and (Stykel, 2004) for linear descriptor systems.The Riccati solution is used to generalize the Linear Quadratic Gaussian balanced truncation (LQGBT) of (Verriest, 1981) and (Jonckheere and Silverman, 1983). These are applied to an electric machine model to reduce the number of states from $>$100000 to 8 while improving accuracy over the state-of-the-art modal truncation of (Zhou, 2015) for the purpose of condition monitoring.Furthermore, a link between unconstrained model predictive control (MPC) with a terminal penalty and LQG of a linear system is noted, suggesting an LQGBT reduced model as a natural model for reduced MPC design.The efficacy of such a reduced controller is demonstrated by the real-time control of a diesel airpath.Model reduction generally introduces modeling errors, and controlling a constrained plant subject to modeling errors falls squarely into robust control.A standard assumption of robust control is that inputs/states/outputs are constrained by convex sets, and these sets are ``tightened;;;; for robust constraint satisfaction.However, robust control is often overly conservative, and resulting control strategies cannot take advantage of the true admissible sets.A new reduction problem is proposed that considers the reduced order model accuracy and constraint conservativeness.A constant tube methodology for reduced order constrained MPC is presented, and the proposed reduced order model is found to decrease the constraint conservativeness of the reduced order MPC law compared to reduced order models obtained by gramian and LQG reductions.For nonlinear systems, a reformulation of the empirical gramians of (Lall et al., 1999) and (Hahn et al., 2003) into simpler, yet more general forms is provided.The modified definitions are used in the balanced truncation of a nonlinear diesel airpath model, and the reduced order model is used to design a reduced MPC law for tracking control.Further exploiting the link between the gramian and Riccati solution for linear systems, the new empirical gramian formulation is extended to obtain empirical Riccati covariance matrices used for closed-loop model order reduction of a nonlinear system. Balanced truncation using the empirical Riccati covariance matrices is demonstrated to result in a closer-to-optimal nonlinear compensator than the previous balanced truncation techniques discussed in the dissertation.