Jacobian-Based Adversarial Attacks for N-1 Robust Power Grid Control

It is crucial to transmit electricity safely and reliably in power grids. This has become increasingly challenging with unstable power generation due to the increased use of renewable energy, unbalanced power consumption, and unexpected events occurring due to natural disasters. Consequently, it is essential to utilize data-driven decision-making methods, such as Reinforcement Learning (RL)-based solutions, for power grid operation. For RL-based solutions to be viable in real-world applications, they usually require robustness against system uncertainty, such as the N-1 safety criteria. Recent efforts to ensure N-1 safety in RL agents for power grid management involve training a separate adversary, which is complex and unstable. In this work, we forgo training an additional adversary and instead, we generate gradient based adversarial attacks on observations. This method enables us to generate long term disturbances by dynamically increasing the complexity of the task at each step, resulting in a faster learning process. We further demonstrate the potential of our approach in producing a robust policy across different grids from Learning to Run a Power Network (L2RPN) for both normal and N-1 contingency scenarios.