The framework of Linear Parameter Varying (LPV) systems has been established with the purpose to represent complex nonlinear and/or time-varying systems in terms of simple linear surrogate models that can easily be used for analyzing the system behavior or designing controllers. In this talk, we give an overview of how LPV surrogate models in a state-space form are currently identified in practice using minimization of the prediction error via quasi-Newton methods, how estimation under noise is handled in these methods, and how initialization of the optimization is accomplished. We will show that how the classical methods have been used to formulate efficient deep neural network-based state-space methods capable for automated model structure selection and efficient choice of scheduling variables. Furthermore, we introduce a novel approach by which general state-space neural network models can be auto-converted to an LPV surrogate form, which can be used to solve nonlinear predictive control problems as successive solutions of linear model predictive problems, corresponding to quadratic programs. We will showcase both the identification and predictive control methods on application examples.
Join at: imt.lu/aula2