Download PDFOpen PDF in browserInvestigating the Effects of Prompt Design on LLM-Generated Priors for Bayesian Calibration in Building Energy Simulation10 pages•Published: June 2, 2026AbstractBayesian calibration has been increasingly used for improving the accuracy of building energy models, but its performance strongly depends on the selection of prior distributions. Large language models (LLMs) provide a new potential way to generate prior distributions by extracting domain knowledge. However, prompt design during LLM execution is critical to the reliability of priors generated by LLMs. The quality and structure of prompts determine how LLMs interpret domain knowledge and translate it into priors, which in turn could influence both prior characteristics and calibration performance. This study investigates how different prompt designs affect the statistical characteristics of LLM-generated priors in building energy simulation calibration. Using three years of monthly electricity data from a campus building, three levels of prompt information (low, medium, and high) were applied to generate priors for four typical parameters. In this study, results show that prompts with lower information levels produced more balanced priors, leading to faster convergence and higher calibration accuracy. These findings suggest that prompt design should balance informativeness and generality to achieve effective LLM-assisted Bayesian calibration, providing a new perspective for integrating LLMs into building energy simulation and modeling.Keyphrases: bayesian calibration, building energy simulation, large language model, prior generation, prompt design In: Wesley Collins, Anthony Perrenoud and John Posillico (editors). Proceedings of Associated Schools of Construction 62nd Annual International Conference, vol 7, pages 445-454.
|
