Data-Driven Model Reduction and Nonlinear Model Predictive Control of an Air Separation Unit by Applied Koopman Theory

Model reduction using Koopman theory as well as the related dynamic mode decomposition (Schmid, 2010), Computationally tractable models are a main requirement build on a lift-and-project concept and aim to construct linear for real-time NMPC (Marquardt, 2002). Data-driven nonintrusive representations of nonlinear dynamics through (nonlinear) model reduction comprises a class of model-free coordinate transformation. Applied Koopman theory has methods for producing low-order representations of highorder a system-theoretic foundation and naturally combines simple dynamical systems from data, e.g., Antoulas et al. dynamic forms with data-driven identification of coordinate (2017). Similar to classical model reduction approaches transformations, e.g., through Kernel methods (Williams (Marquardt, 2002), these data-driven methods project a highorder et al., 2015), deep learning (Lusch et al., 2018), or sparse regression system from the full state space to a lower dimensional techniques (Brunton et al., 2016).