Florian Dörfler: Online Feedback Optimization
in
Workshop: Foundations of Reinforcement Learning and Control: Connections and Perspectives
Abstract
Online feedback optimization refers to a class of feedback controllers that take the form of online optimization algorithms, that are real-time interconnected with a physical system, and that asymptotically steer the system to the solution of an optimization problem while respecting physical and operational constraints. In comparison to other optimization-based control strategies, transient optimality of trajectories is not the primary goal, and no predictive model, running costs or exogenous set-points are required. Hence, one aims at controllers that require little (respectively no) model information, demand low computational cost, but that leverage real-time measurements to stabilize the closed-loop system while enforcing constraints and seeking asymptotic optimality. Online feedback optimization method sits square between optimal/adaptive control and reinforcement learning, it has historic roots in communication networks and process control, and our enabling application of interest is the electric power system. I will give a tutorial introduction, discuss different algorithms, their closed-loop stability when interconnected with physical systems, and implementation aspects. Throughout the talk I demonstrate the potential of our methodology for real-time operation of power systems including commercial implementations.
Speaker
