Integrating Large Language Models and Reinforcement Learning for Efficient Home Energy Management

Abstract

The home energy management system (HEMS) has gained significant attention with the advancement of smart monitoring and Internet of Things (IoT) technologies. Data-driven approaches, particularly reinforcement learning (RL), have shown promise in learning HEMS scheduling policies through environment interaction but often suffer from poor sample efficiency and high computational overhead in environments with delayed rewards. To address these challenges, this article proposes a hybrid framework integrating large language models (LLMs) and RL for efficient and safe HEMS scheduling. Unlike existing LLM-based energy studies that mainly use LLMs for forecasting, knowledge processing, or open-loop advisory support, the proposed framework deploys an LLM as a direct decision-making actor in a closed-loop HEMS control setting. By leveraging the LLM’s pretrained reasoning capability as an implicit policy prior, the proposed method improves the sample efficiency of RL-based HEMS scheduling without requiring expert demonstrations. Proximal policy optimization (PPO)-based fine-tuning with low-rank adaptation (LoRA) further enables efficient LLM–HEMS alignment by updating only approximately 0.1% of the model parameters. Moreover, a constrained action-sampling mechanism restricts LLM outputs to predefined admissible appliance actions, thus ensuring physically executable, constraint-compliant scheduling decisions. Experiments driven by real-world market and appliance data show that the proposed method improves home utility by 33.3%, reduces user discomfort costs by 28.6%, and requires over 20 times fewer training episodes than conventional RL baselines.