Open Collections

Rostam, Mohammadreza

University of British Columbia

As the computational power available to control engineers continues to grow, Model Predictive Control (MPC) systems are increasingly recognized for their potential to revolutionize various industrial and technological domains. These advanced control systems offer the promise of optimal control inputs while inherently addressing system constraints. However, their effectiveness hinges on the accuracy of their predictions of future system behavior, a challenge often exacerbated by the presence of external disturbances. Inaccurate handling of these disturbances can lead to a scenario where the performance of a computationally expensive MPC system is worse than that of a simpler, more traditional Proportional-Integral-Derivative (PID) controller. This thesis explores the potential of Gaussian Process (GP) models, a well-established machine learning technique renowned for its inherent ability to manage uncertainty, to enhance the prediction component of MPC systems. Specifically, the research aims to develop a framework for integrating GP models into MPC systems to bolster their robustness against uncertainties arising from external disturbances. The research commences with a comprehensive framework designed for systems that require long prediction horizons. This framework incorporates a novel method for effectively utilizing GP models in scenarios characterized by significant uncertainty in disturbances. The research then refines this framework to customize it for situations involving quasi-periodic disturbances, a common phenomenon in energy systems where user behavior exhibits unknown and variable patterns. To validate the efficacy of the proposed control system, a series of simulations are conducted on a domestic solar thermal system. The results of these simulations reveal that the proposed system can achieve performance within 2.5% of a hypothetical scenario where perfect future knowledge is available. Moreover, the system demonstrates a threefold performance improvement compared to the simple PID controller, a widely used control strategy in industry for such systems. Additionally, the thesis explores the theoretical aspects of the proposed system, including stability and recursive feasibility, when applied to simple linear models using a Gaussian Process (GP) model for prediction. This analysis provides valuable information on the theoretical foundations and practical implications of the proposed approach.

Text

2024-11-12

Attribution-NonCommercial-NoDerivatives 4.0 International

http://hdl.handle.net/2429/89648

Mechanical Engineering

University of British Columbia

UBCV

http://creativecommons.org/licenses/by-nc-nd/4.0/