Model Predictive Control can be used to optimize industrial processes such that they maintain high production efficiency whilst staying within their operational constraints. This thesis presents an end-to-end data-driven method for implementing Model Predictive Control on a well-known industrial process control simulator using both Linear and Neural Network Models.
Implementing the controller does not require a detailed understanding of the underlying process. Particular methods for estimating the quality of collected data and models are presented and a comparison between Linear and Neural Network Models is made. The process of implementing Model Predictive Control using these two models and gradient based planning is presented, along with methods to address their shortcomings. A series of quantitative benchmarks is presented showing the performance of each model during Model Predictive Control in a variety of different situations and the reasons behind differing performance are discussed.
