Blackout Mitigation Via Physics-guided RL

This paper considers the sequential design of remedial control actions inresponse to system anomalies for the ultimate objective of preventingblackouts. A physics-guided reinforcement learning (RL) framework is designedto identify effective sequences of real-time remedial look-ahead decisionsaccounting for the long-term impact on the system's stability. The paperconsiders a space of control actions that involve both discrete-valuedtransmission line-switching decisions (line reconnections and removals) andcontinuous-valued generator adjustments. To identify an effective blackoutmitigation policy, a physics-guided approach is designed that uses power-flowsensitivity factors associated with the power transmission network to guide theRL exploration during agent training. Comprehensive empirical evaluations usingthe open-source Grid2Op platform demonstrate the notable advantages ofincorporating physical signals into RL decisions, establishing the gains of theproposed physics-guided approach compared to its black box counterparts. Oneimportant observation is that strategically removing transmission lines,in conjunction with multiple real-time generator adjustments, often renderseffective long-term decisions that are likely to prevent or delay blackouts.