BENCHMARKING LARGE LANGUAGE MODELS (LLMS) FOR DATA‑DRIVEN END‑USE ENERGY ANALYSIS IN U.S. RESIDENTIAL BUILDINGS

Occupant behavior significantly influences residential energy consumption, yet traditional energy modeling practices commonly rely on fixed activity schedules, neglecting dynamic behavioral variations. Among household activities, Dishwashing is a notable yet often overlooked contributor to residential energy demand, as it involves both electricity and hot-water use. Its frequent occurrence after dinner significantly adds to evening peak loads. To address this gap, this study develops a high-resolution, data-driven framework specifically focused on modeling residential dishwashing behavior to support demand-flexible energy management strategies. Utilizing detailed temporal data on dishwashing activities extracted from the American Time Use Survey (ATUS), complemented by relevant household and temporal covariates, this research applies advanced machine-learning algorithms to predict both the probability and timing of dishwasher operation. The resulting occupant-informed load profiles are integrated into comprehensive building energy simulation models, facilitating the assessment of peak-shifting potential, energy-efficiency improvements, and demand-response effectiveness. Findings from this analysis provide enhanced predictive accuracy for energy demand, inform the development of occupant-centered appliance control strategies, and yield actionable insights and recommendations for incentive design and retrofit policies. Furthermore, the proposed modeling framework offers flexibility for adaptation to other occupant-driven activities, promoting broader scalability in occupant-centric energy modeling and supporting the transition toward resilient, low-carbon energy systems.