Physics-Grounded Deep Reinforcement Learning for Adaptive Out-of-Step Protection in Low-Inertia Power Grids

This NSF project aims to develop a physics-grounded artificial intelligence (AI) framework for power grid protection that keeps grids safe as they undergo rapid and fundamental change. A critical threat to grid reliability is the out-of-step (OOS) condition, where the loss of synchronization can cause permanent equipment damage and trigger cascading blackouts. As inverter-based resources grow and system inertia declines, OOS conditions become more likely and far harder for existing protective relays to handle. The project will bring transformative changes to OOS protection by enabling relays to continuously adapt their behavior based on real-time grid conditions rather than predetermined fixed parameters. This will be achieved by combining physics-based analysis of grid dynamics with machine learning in a way that keeps protection decisions transparent and verifiable by engineers. The intellectual merit of the project includes establishing scientific foundations for integrating physical domain knowledge into machine learning frameworks under real-time and regulatory constraints, advancing fundamental understanding of trustworthy AI in safety-critical infrastructure. The broader impacts of the project include enhancing national grid reliability and resilience, training the next generation of power system engineers through graduate research and an NSF-funded undergraduate research program, and engaging K-12 students in electrical engineering through community outreach. This project investigates a physics-grounded deep reinforcement learning (DRL) framework that preserves the industry-standard OOS protection architecture while making its settings continuously adaptive. A key challenge is that today's relays depend on fixed settings derived from labor-intensive engineering studies, and these settings cannot keep pace with continuously shifting grid conditions. Dynamic state estimation (DSE) addresses this by reconstructing the unmeasurable states and identifying the center of oscillation in real time, providing physically meaningful features for DRL control and cross-system generalization. The DRL agent uses these features to continuously optimize relay settings, while all protection decisions are executed by the classical relay logic to preserve engineering interpretability and compliance with industry reliability standards. A meta-learning framework further enables the learned policy to adapt rapidly to new network topologies and operating conditions from limited experience, reducing the engineering effort required for deployment on new utility systems. The framework will be validated across multiple test environments of increasing complexity. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.