Data Enhanced Model Predictive Control of a Coupled Tank System

Many advanced methods of process control are based on underlying process models. Therefore, the control performance usually depends strongly on the model accuracy, e.g. the extent of deviations between the real plant behavior and its mathematical representation. Such an advanced method is model predictive control, which has been successfully established in various fields. A model predictive control is based on a sufficiently accurate model of the system to optimize performance and ensure constraints are met. In practice, model descriptions can be subject to significant inaccuracies due to, e.g. insufficient data, unknown dynamics, or system changes. In addition, most processes in industry show time variant dynamics, e.g. due to wear and tear. Modeling such varying characteristics is usually impractical or requires a disproportional amount of modeling effort. Furthermore, the resulting models are usually too complex for real-time process control. In this work, a model predictive control based on a grey-box model is presented. While the model’s basic structure is given by a linearized physical model it is enhanced by a data-driven part using a long short-term memory network. An approach to formulate the linear, time-varying, model predictive control problem that takes into account the model uncertainties of the nominal model predicted by the long short-term memory network model is described. The improvements in control performance are demonstrated on a real demonstrator, which is a coupled tank system. The approach presented reduces the average absolute error by 69.16%, and overshooting by 11.7% on average.