Integrating Large Language Models with Optimization for Interactive Power Systems Planning

Optimization models can inform decision-making for infrastructure planning, but interacting with these models often requires knowledge of model implementation, data structures, and mathematical programming software. We present a natural-language interface based on the Model Context Protocol for power grid capacity expansion planning under operational uncertainty. The interface enables users to inspect model data, modify planning scenarios, solve optimization problems, and query results without directly interacting with the underlying codebase. We evaluate the interface using Claude Opus 5 and compare identical planning tasks against direct large language model access to the capacity expansion planning codebase. Both approaches completed the evaluated tasks correctly, but direct codebase access required approximately 81% more tokens across a set of six representative tasks that users might request. These results show that a structured tool interface can reduce token usage while providing controlled access to optimization models.