Structure Exploiting Reinforcement Learning Control with Applications to Power System Dynamics

Presenters: Sayak Mukherjee, NC State University & FREEDM Systems Center PhD Candidate and Aranya Chakrabortty, PhD, Associate Professor NC State University

Designers of dynamic models for controlling large-scale dynamic systems like power networks may not have all the necessary information. Reinforcement Learning (RL) can be used to design feedback control from state trajectory measurements without the explicit model information. However, full dimensional RL designs for large-scale systems are limited by the curse of dimensionality that increases learning time and leads to dense feedback control structures. In this webinar, we will incorporate the ideas of model reductions to perform a reduced-dimensional learning control. We will first explore the measurement-driven design for power systems on a transmission-scale model of the New York State grid by highlighting an indirect system identification optimal control method. We will also present a brief overview of the RL control approach that exploits the structure of dynamic systems with two time-scale behaviors and translate this for interconnected power grids with clusters. Our results show that the control design improves the dynamic performance and reduces inter-area oscillations. Lastly, we will discuss a decentralized RL control variant for distributed energy resources (DERs) that shows modularity and scalability for futuristic power systems.

Register HERE