System identification for MPC

One of the main challenges of model predictive control (MPC) is the requirement of a dynamic system model. Such a model can be obtained from first principles, typically requiring significant time and engineering effort. Alternatively, a data-based approach can be persued, if system measurements are available. System identification has recently regained attention but has been an ongoing field of research in control for many years. Its importantance is becoming more pronounced in todays world that provides data at an increasing rate and requires smarter control decisions, to solve our current and future environmental and societal challenges.

Model predictive control allows for a number of new development in the traditional field of system identification. In particular, its ability to handle non-linear system models makes it possible to use a number of powerful non-linear identification techniques. Deep learning-based surrogate models for MPC are a promising application that we are actively researching at the Laboratory of Process Automation Systems.

Model predictive control is also special in the sense that predictions of finite sequences of the future system behavior are required. These sequences can be obtained recursively from classical system models that predict the next state of the system. On the other hand, MPC can profit from models that directly provide such a sequence. The Laboratory of Process Automation Systems researches in particular the popular data-enabled predictive control (DeePC) approach and the closely related subspace predictive control (SPC).

Coming soon.