PowerAtlas: Feasibility-Aware Language Models for Electricity-Computing Co-Scheduling
The paper addresses the urgent need to coordinate electricity and computing schedules for growing AI workloads by treating it as a single decision problem for a Large Language Model (LLM). It proposes PowerAtlas, a framework that uses supervised learning followed by Feasibility-Aware Group Relative Policy Optimization (FA-GRPO) to generate joint decisions. This allows the agent to produce schedules that are both physically feasible and economically efficient.