Agentic AI Home Energy Management System: Residential Load Scheduling (arxiv.org)

Researchers introduced an agentic AI Home Energy Management System (HEMS) that uses large language models as autonomous coordinators to translate natural-language household requests into multi-appliance schedules and device control. The system uses a hierarchical agent architecture—one orchestrator plus three specialist agents—employing the ReAct pattern for iterative reasoning and dynamic coordination (no hardcoded workflows or example demonstrations needed). It also integrates Google Calendar to extract context-aware deadlines. Evaluated on real Austrian day-ahead electricity prices, the setup benchmarks schedules against a mixed-integer linear program (MILP); Llama-3.3-70B coordinated all appliances across scenarios and matched MILP cost-optimal results, while other models achieved perfect single-appliance scheduling but failed at full multi-appliance coordination. Significance for AI/ML: this work demonstrates that LLMs can act beyond code generation or parameter extraction to autonomously run end-to-end control workflows in a safety- and cost-sensitive domain, lowering the usability barrier for residential demand response. Key technical implications are that model choice and prompting critically affect multi-agent coordination performance, analytical query handling remains unreliable without explicit guidance, and hierarchical agent design plus external context tools (calendar, price feeds) enable practical, explainable scheduling. The authors open-source the orchestration logic, prompts, tools, and web UI to spur reproducibility and further research into agentic control for energy systems.