Direct learning of LPV controllers from data

In many control applications, it is attractive to describe nonlinear (NL) and time-varying (TV) plants by linear parameter-varying (LPV) models and design controllers based on such representations to regulate the behavior of the system. The LPV system class offers the representation of NL and TV phenomena as a linear dynamic relationship between input and output signals, which is dependent on some measurable signals, e.g., operating conditions, often called as scheduling variables. For such models, powerful control synthesis tools are available, but the way how to systematically convert available first principles models to LPV descriptions of the plant, to efficiently identify LPV models for control from data and to understand how modeling errors affect the control performance are still subject of undergoing research. Therefore, it is attractive to synthesize the controller directly from data without the need of modeling the plant and addressing the underlying difficulties. Hence, in this paper, a novel data-driven synthesis scheme is proposed in a stochastic framework to provide a practically applicable solution for synthesizing LPV controllers directly from data. Both the cases of fixed order controller tuning and controller structure learning are discussed and two different design approaches are provided. The effectiveness of the proposed methods is also illustrated by means of an academic example and a real application based simulation case study.