Large language model-driven rule synthesis for adaptive home energy management

This paper proposes an innovative large language model-driven rule synthesis (LLM-RS) framework for adaptive home energy management system (HEMS). The framework addresses three critical challenges in existing HEMS: interpreting natural language user intentions, accommodating diverse household configurations, and ensuring transparent and deployable control rules. The LLM-RS framework consists of three synergistic modules: semantic intent extraction using an LLM-based parser, automated translation of structured intents into executable control logic through a rule generation engine, and seamless deployment within an existing rule-based control infrastructure. The system enables users to express control preferences through flexible natural language instructions, which are semantically parsed and automatically converted into IF-THEN rules without requiring manual configuration or pre-collected data. Experimental results with both open-source and proprietary LLMs demonstrate that GPT-4o achieves the highest performance across intent accuracy, slot alignment, and rule match accuracy metrics, while smaller models show significant degradation. The LLM-RS approach represents a paradigm shift from traditional rigid rule-based or complex optimization-based HEMS toward user-centric, plug-and-play energy management, enabling rapid deployment across heterogeneous household configurations while maintaining transparency and interpretability of control decisions.