Physics-Informed Deep Reinforcement Learning for Inter-Area Oscillation Damping via Shapley-Based Generator Selection

Inter-area oscillations (IAOs) pose significant stability challenges in modern power grids, especially under topology changes such as line outages. Traditional methods lack adaptability, while artificial intelligence(AI)-driven approaches often struggle with interpretability, safety, and generalization. This paper proposes a physics-informed deep reinforcement learning (DRL) framework for wide-area damping control, featuring: (1) eigenvalue optimization in a Markov Decision Process (MDP), (2) SHapley Additive exPlanations (SHAP) for interpretable and cost-efficient generator selection, and (3) graph embeddings for topology-aware adaptability. A safety envelope governed by small-signal stability criteria improves robustness by discouraging unstable actions. Validation on the IEEE 39-bus system shows scalability, resilience to contingencies without retraining, and compliance with safety constraints, making it a practical solution for renewable-rich power grids.