RestoreBench Evaluates AI Agents for Power Grid Restoration

Large Language Model (LLM) agents are increasingly being used to automate complex, multi-step engineering workflows, leveraging their ability to use tools, interpret intermediate results, and plan iteratively. A promising, yet underexplored, application area for these agents is the diagnosis and resolution of non-convergent power flow scenarios within electrical grids. This task demands significant engineering judgment, experimental approaches, and precise decision-making within predefined action spaces. To facilitate research in this domain, a new benchmark called RestoreBench has been introduced. This benchmark specifically evaluates the capabilities of LLM agents in restoring power flow convergence. It assesses different LLM architectures, including chatbot, single-agent, and multi-agent systems. The evaluation covers two distinct power grids, each presenting 46 unique cases that require one or more corrective actions to achieve convergence. RestoreBench clearly defines the simulation environment, the available observation and action spaces, and the metrics for evaluation, thereby establishing a reproducible framework for the development of agentic AI systems tailored for power system planning and operational tasks.