When
Where
Speaker: Robert Ferrando, Program in Applied Mathematics
Title: Physics-Informed Reinforcement Learning for Emergency Fuel Transition
Abstract: Consider a power grid consisting of generators primarily powered by natural gas. We study the operational task of responding to an emergency scenario precipitated by a sudden interruption to the gas supply. It is possible to transition units to a secondary fuel source, such as diesel, or to shut them off. The decision-making process is inherently challenging, as there is a need to rapidly balance the interests of the system operator -- to minimize cost -- and of customers -- to maximize supply. As well, the power and gas systems operate on different time scales, and modeling the gas system with high fidelity requires the efficient numerical solution of a system of PDEs. We propose a Markov Decision Process (MDP) framework to model the transitions, and compare the performance of several policies guided by engineering intuition. Finally, we discuss how to leverage deep reinforcement learning to iteratively recover an optimal policy. This problem is of interest to many small, relatively isolated power grids both inside the US and internationally.