PowerAtlas Optimizes Electricity-Computing Co-Scheduling with LLM Agents.

The escalating demand from AI workloads is transforming data centers into substantial, dynamic, and geographically flexible loads on power grids, creating an urgent need for synchronized electricity and computing scheduling. Traditional scheduling methods, especially those from general-purpose large language models (LLMs), often produce infeasible schedules that can lead to grid violations and unserved power. PowerAtlas addresses this by introducing an LLM-agent framework for electricity-computing co-scheduling. This framework intelligently combines historical operational data, expert domain knowledge, and strict physical constraints to generate joint decisions. These decisions are designed to satisfy both the operational rules of the power grid and the service-level agreements (SLAs) of computing tasks. Developed in collaboration with a Chinese provincial power utility, PowerAtlas was validated on an experimental electricity-computing network using real data center information. The project also created ECBench, a benchmark dataset of 2,000 scheduling instances with optimal solutions derived from de-identified operational data. Experiments across various LLMs confirm PowerAtlas's effectiveness in realistic conditions, showing consistent feasibility and cost benefits with open-weight LLM backbones.