AlphaZero-Inspired AI Stabilizes Power Grids with Topological Control

Modern power grids face increasing strain due to the integration of volatile renewable energy sources. Reinforcement Learning (RL) offers a promising avenue for autonomous topological reconfiguration, a cost-effective method to manage grid congestion compared to traditional redispatching. However, the vast combinatorial action space and strict operational constraints pose significant challenges. This research explores the effectiveness of model-based AlphaZero-inspired approaches, which leverage Monte Carlo Tree Search (MCTS), for proactive grid management. The study systematically evaluates how different reward functions, observation densities, and search guidance mechanisms impact an agent's ability to maintain grid stability. The findings show that an optimized AlphaZero approach achieves a peak survivability of 98.43%, significantly outperforming a Proximal Policy Optimization (PPO) variant. Key insights include that MCTS can be more efficient without guidance from a prior learned policy or value function, and that simple binary survival rewards are more effective than complex multi-objective functions. The research concludes that while AlphaZero is powerful, effective and reliable grid control requires a "minimalist" integration of domain-specific heuristics, binary rewards, and a restricted observation space.