Evaluation of Data Enhanced Model Predictive Control for a Coupled Tank System

IEEE Conference on Control Technology and Applications (CCTA), vol. 3, pp. 79–84

Abstract

Model predictive control (MPC) depends on a model of a process to predict the future behavior. These predictions are used to optimize control actions. Grey-box models combine first principle knowledge of the process and knowledge based on process data. They are useful for MPC because they can be simpler to develop and adapt than complex white-box models. Yet these models are, unlike full black-box models, for the most part interpretable. This study compares the performance of different grey-box models in MPC, applied to a coupled tank system. The performance is evaluated based on the ability to accurately predict future outputs and optimize control actions. In this regard, Gaussian process regression, long short-term memory networks and gated recurrent unit networks are examined in more detail. Overall, this study demonstrates the potential of grey-box models for MPC and highlights the importance of carefully selecting the appropriate model for a given process. The results show that the performance and computational effort varies depending on the specific model used in the control system. However, the grey-box models are able to improve the performance of the control system compared to using only a white-box model up to 96.1% regarding the steady-state behavior.

Authors 3

1. Dominik Scheurenberg Aachen Institute of Automatic Control (IRT)

   RWTH Aachen University

   Affiliation as printed

   RWTH Aachen University,Institute of Automatic Control,Aachen,Germany,52062

2. Sebastian Stemmler Aachen Institute of Automatic Control (IRT)

   RWTH Aachen University

   Affiliation as printed

   RWTH Aachen University,Institute of Automatic Control,Aachen,Germany,52062

3. Dirk Abel Aachen Institute of Automatic Control (IRT)

   RWTH Aachen University

   Affiliation as printed

   RWTH Aachen University,Institute of Automatic Control,Aachen,Germany,52062