Harnessing Agent Capabilities to Integrate Human Demands into Energy Infrastructure Operations

Abstract

Energy infrastructure systems, e.g. EV charging networks, smart buildings, must make real-time decisions to satisfy time-varying, context-dependent human demand. Yet existing methods fall short: edge controllers cannot interpret human-demand contexts, whereas LLM agents that excel at fusing such human information are unreliable and latencyprone for direct physical control. We propose Human-Aware Agentic Operation (HAAO), a hybrid framework that involves these two reasoning modes: at the upper level, in-context prompting leverages a black-box LLM agent to fuse community context and predict human demand in a structured, machine-readable form; at the lower level, a lightweight model maps this human context and system states to physical control actions. The performance of HAAO is evaluated on the EV charging power management problem.