LICA: A multi-agent framework for household electricity scheduling under real-time pricing

Real-time electricity pricing enables households to lower bills, yet effective rescheduling requires private and user-specific preferences that are rarely known a priori. This paper develops an intelligent electricity-consumption advisor that learns latent user preferences through natural-language interaction and generates personalized, comfort-aware schedules under real-time pricing. The key challenge lies in jointly managing uncertain preferences, user interaction cost, and downstream scheduling quality. We address this by framing preference acquisition as a sequential decision-making problem that balances preference-estimation accuracy and interaction burden, and by introducing LICA, a multi-agent LLM-based framework that integrates adaptive querying with preference-aware optimization. A bounded-depth tree search provides an oracle benchmark, while scalable approximate policies enable efficient real-time deployment. Experiments across synthetic and real-world settings show that LICA consistently improves scheduling quality while significantly reducing interaction burden. On real Finland price traces, LICA lowers realized electricity costs by approximately 12% compared with a single-turn LLM baseline while fully meeting departure–energy requirements, demonstrating practical effectiveness for real deployments.