19 August 2025 in Neurocomputing
7 October 2025 in IEEE Internet of Things Journal
The global imperative for carbon neutrality drives unprecedented transformation in building energy systems, where battery energy storage systems (BESS) integrated with photovoltaic (PV) installations offer substantial demand-side flexibility. However, existing battery scheduling approaches—including heuristic rules, mathematical optimization, and reinforcement learning (RL)—fundamentally cannot incorporate user-informed future events lying outside historical training distributions, limiting their ability to leverage human foresight for proactive energy management. This work introduces a human-in-the-loop battery scheduling mechanism via intent-guided rule-RL that systematically bridges user cognitive foresight with autonomous control systems. The mechanism employs large language models (LLMs) to extract structured battery preconditioning intents from natural language inputs through interactive dialog, implements heuristic rules for time-bounded state preparation, utilizes soft actor-critic (SAC) RL for baseline control, and coordinates these paradigms through hierarchical priority arbitration. Experimental validation using real residential energy data demonstrates substantial performance improvements: 39.2% cost reduction in discharge scenarios, 15.1% cost reduction in charge scenarios, and effective multi-intent coordination achieving 6.3% cost reduction with 20.0% improvement in cost savings compared to pure RL approaches. This research establishes a new paradigm for building energy management where human knowledge becomes a strategic optimization resource.
1 May 2026 in Cyber-Physical Energy Systems