Toward An Explainable Electric Power Grid Operation Assistant Using Large Language Models

Abstract

This thesis explores potential applications of LLMs for assisting the analyses and decisionmaking of complex electric power grid operators. The power grid is a critical piece of infrastructure currently challenged by increased electrification, integration of renewable energy sources, and distributed energy resources (DERs). Human operators struggle to process the massive amounts of data produced by modern smart grids and need innovative solutions to handle the increased complexity of operational decisions. This thesis investigates the potential role of Large Language Models (LLMs) in grid operation tasks, focusing on interpretability and generalizability while exploring how LLMs can assist operators by providing actionable insights and recommendations. Multiple versions of LLM agents were developed, including naive and tool-assisted designs, and were evaluated on the Learn to Run a Power Network (L2RPN) benchmark for steady-state and cascading failure scenarios. While the LLM agents performed better in scenarios requiring exploratory decision-making, they struggled in steady-state operation and were constrained by their integration with tools and the testing environment. This work was limited by compute constraints, which affected the choice of model and the length of evaluation scenarios, and future work is needed toward seamless interaction of LLMs and power systems simulators, however LLMs have the potential to transform future grid operation, paving the way for more resilient and sustainable energy sector of the 21st century.