EXPLAINABLE PARAMETER CALIBRATION VIA IMPORTANCE-DRIVEN SEQUENTIAL DESIGN WITH AN APPLICATION TO BUILDING ENERGY SYSTEMS

• 成果类型: Article
• 署名作者: Jeong, Cheoljoon; Byon, Eunshin
• 署名单位: Clemson University; University of Michigan System; University of Michigan
• 刊物名称: ANNALS OF APPLIED STATISTICS
• ISSN/ISSBN: 1932-6157; 1941-7330
• DOI: 10.1214/26-AOAS2148
• 发表日期: 2026-06
• 页码: 1562-1585
• 关键词: Key words and phrases. Bayesian optimization; efficient global optimization; Multi-Armed Bandit; sequential design; GLOBAL SENSITIVITY-ANALYSIS; bayesian calibration; models; optimization; FRAMEWORK

摘要

Parameter calibration seeks to estimate unobservable parameters in a computer model by aligning field observations with computer model outputs. In the building energy sector, a physics-based computer model is developed to analyze building energy use, given various weather conditions and operational scenarios. To obtain accurate simulations, it is necessary to calibrate model parameters required for preconfiguration. Among various techniques, Bayesian optimization stands out for its potential but faces some challenges when handling a large number of parameters. A possible remedy is to focus selectively on influential parameters, thereby simplifying a complex, high-dimensional task into a more tractable, lower-dimensional endeavor. We develop a new method that ranks parameter importance to effectively enable stochastic dimension reduction by utilizing the multi-armed bandit approach. By accounting for unequal importance among parameters, our approach generates accurate surrogate models tailored to the reduced dimension and guides an efficient exploration of the parameter search space in the Bayesian optimization procedure. The numerical studies and building energy simulation case study demonstrate that the proposed approach achieves a significant improvement in both calibration accuracy and efficiency. Moreover, it capably identifies the influential parameters and explains their impact on the computer model, providing valuable insights into understanding the system's dynamics.