From Static Surveys to Living Agents: A Generative LLM-Agent Workflow for Residential Energy Simulation

Occupant behavior fundamentally drives urban energy consumption, yet it remains highly stochastic and challenging to model. Traditional Agent-Based Models (ABMs) rely on rigid, hard-coded rules that fail to capture the nuanced and dynamic nature of human daily life. While Large Language Models (LLMs) offer promising potential for generative social simulation, ensuring robust personality profiling and behavioral consistency—both ex-ante formulation and ex-post verification—remains largely unexplored. To bridge this gap, this paper presents a novel workflow that transforms household questionnaire responses into simulation-ready agent profiles. By leveraging AgentSociety, a large-scale generative-agent simulation framework, we execute multi-day routines and map the resulting outputs to demand-relevant occupancy and activity traces. Using a private household energy survey from the Lausanne area, we simulate separate 48-hour workday and weekend scenarios for 500 agents. We comprehensively evaluate the workflow along three practical axes: contextual differentiation across day types, alignment with survey-derived behavioral anchors, and plausibility of aggregate occupancy and activity patterns. Results show clear workday–weekend distinctions and remain consistent with daily occupancy patterns reflected in both surveys and local standards. At the population level, the simulated traces reproduce structured occupancy rhythms, heterogeneous occupancy regimes, and non-trivial activity timing patterns. Ultimately, these findings establish a foundational framework to incorporate household survey data into executable LLM-driven behavioral simulation. By coupling generative social dynamics with physical energy models, this framework sets the stage for future human-centric energy optimization and the comprehensive assessment of demand-side flexibility in smart grids.