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

Buildings consume approximately 40% of national energy in the U.S., significantly contributing to greenhouse gas emissions. Data-driven analysis and accurate prediction of energy consumption is vital for advancing sustainability and climate objectives. In this regard, Large Language Models (LLMs), advanced AI systems trained on vast datasets to process and generate human-like text, are pivotal in enhancing data-driven analysis and prediction of building energy use. With a variety of LLMs available, selecting the most effective model is critical for optimizing energy consumption forecasts. This study investigates, analyzes, and benchmarks multiple LLMs to identify the optimal model for data-driven analysis of energy consumption in U.S. residential buildings, leveraging data from the Residential Energy Consumption Survey (RECS). Through rigorous evaluation of model performance across different end-use energy levels—including space heating, air conditioning, water heating, lighting, and appliances—this research identifies the most accurate and efficient LLM. The identification of the best-performing LLM informs retrofit planning, energy policy development, and demand-side management, enabling more effective energy-saving strategies.