LLM-BEM: Zero-shot LLM-Agent Scheduling for Smart Building Energy Management

Smart buildings require efficient energy scheduling across subsystems to balance cost, comfort, and sustainability. Existing methods lack flexibility, generalization ability, and interpretability under dynamic conditions. We propose LLM-BEM, Zero-shot LLM-Agent Scheduling for Smart Building Energy Management, a zero-shot multi-agent framework that integrates large language models for building energy control. LLM-BEM includes ZoneAgent for local intent generation, BuildingAgent for global reconciliation, and LLM-Scheduler for final decisionmaking and explanation. Simulation results show that LLM-BEM reduces energy costs, defers non-critical loads, and preserves occupant comfort. This work demonstrates the feasibility of prompt-driven, training-free scheduling for generalizable and interpretable smart building management.