Large language model-enhanced home energy management with dynamic user preference elicitation and hierarchical data-sharing

The residential sector has received great attention worldwide due to its significant contribution to energy consumption and carbon emissions. Although Deep Reinforcement Learning (DRL) based Home Energy Management (HEM) has provided promising solutions for user comfort satisfaction and energy savings, existing approaches still face three challenges: (1) accurately perceiving dynamic user preferences, (2) efficient learning with limited data samples, and (3) protecting user privacy during data sharing. To address these issues, a Large Language Model (LLM)-enhanced HEM with dynamic user preference elicitation and hierarchical data-sharing is proposed, introducing three key innovations. First, the Knowledge-Guided Chain-of-Thought method is proposed to perceive dynamic user preferences from unconstrained natural language interactions, by leveraging users’ historical knowledge as explicit guidance. Second, the LLM-generated reference actions based on shared knowledge and preference descriptions are proposed to accelerate the learning process under conditions of cold-start situations. Third, a hierarchical data-sharing mechanism integrated with LLM-based knowledge extraction is introduced to enable effective knowledge inference from natural language while preserving user privacy. Simulation results in the household scenario with dynamic user preferences demonstrate that the proposed framework achieves significant improvements in both user preference elicitation and learning efficiency, ultimately leading to a 28.8% reduction in cost and a 9.3% improvement in comfort compared to fixed preference DRL methods.