Cross-domain indoor performance prediction via in-context learning
A CFD-TabPFN-SHAP framework integrating air quality, energy, and thermal comfort
Xueren Li (China University of Mining and Technology, Jiangsu Engineering Research Center of Dust Control and Occupational Protection, Royal Melbourne Institute of Technology University)
Weijie Sun (University of Alberta)
Liwei Zhang (Purdue University)
Bichen Shang (Purdue University)
Ruipeng Xu (TU Delft - Applied Sciences, J.M. Burgerscentrum Research School for Fluid Mechanics)
Jiyuan Tu (Royal Melbourne Institute of Technology University)
Yuanfu Pei (Gansu Beautiful Technology Industrial Group, Co. Ltd.)
Xiang Fang (Shanghai University of Engineering Science)
More Info
expand_more
Other than for strictly personal use, it is not permitted to download, forward or distribute the text or part of it, without the consent of the author(s) and/or copyright holder(s), unless the work is under an open content license such as Creative Commons.
Abstract
Efficient indoor environmental management is vital for reducing energy use while safeguarding indoor air quality and occupants’ comfort. However, due to difficulty of obtaining comprehensive datasets across multiple domains, many current studies still focus on a single performance metric, with dataset limitations preventing the development of robust and accurate multi-objective predictive models and operate as opaque data-driven systems with limited explanatory capability. This work developed CFD datasets derived from an experimentally validated model to systematically generate data covering parameters of indoor air quality, energy use, and thermal sensation with varied ventilation conditions. This dataset is further used for machine learning (ML) analysis. A cutting-edge Tabular Prior-data Fitted Network (TabPFN) is adopted for multi-target prediction and benchmarked against XGBoost, CatBoost, and Backpropagation Neural Networks (BPNN). SHapley Additive exPlanations (SHAP) method was used to elucidate its predictions, identifying how the most influential parameters govern the variations in indoor environmental behavior. The findings indicate that TabPFN outperforms the other three models in both predictive accuracy and computational efficiency, with MAE reduced by 62.08-95.4 % and RMSE by 49.15-85.50 %, while inference is accelerated by 67.5-85.9 %. SHAP analysis quantified nonlinear and directional contributions of features, linking model outputs to physical mechanisms such as draft risk, thermal sensation, and cooling load demand. The proposed TabPFN-SHAP framework is expected to offer valuable insights to guide the optimisation of building environmental control strategies.
help_outline