Politecnico di Milano
Wednesday, July 18, 2018, 11:00
Recent research efforts in Model Predictive Control (MPC) have been focusing on the identification, or learning, of the system dynamics from experimental data. This trend is motivated by the growing availability of data in real-world applications, and by a gap in the literature between MPC design and model identification for predictive control. In this talk, a comprehensive approach addressing identification and control to learning-based MPC for linear systems is presented. The design technique yields a suitable MPC law, based on a dataset collected from the working plant. The method is indirect, i.e. it relies on a model learning phase and a model-based control design one, devised in an integrated manner. In the model learning phase, Set Membership techniques are used to derive multi-step prediction models for the system, together with the associated uncertainty. In the control design phase, a robust MPC law is proposed, able to track piece-wise constant reference signals, with guaranteed recursive feasibility and convergence properties. The controller embeds the multi-step predictors in the cost function, it ensures robust constraints satisfaction thanks to the learnt uncertainty model, and it can deal with possibly unfeasible reference values.