Predicting wind-driven rain amount on building facade based on machine learning models

Biao Lu, Yunjie Li, Xianqi Huang, Min Jiang, Chi Feng,
Predicting wind-driven rain amount on building facade based on machine learning models,
Building and Environment,
Volume 282,
2025,
113308,
ISSN 0360-1323,
https://doi.org/10.1016/j.buildenv.2025.113308.
(https://www.sciencedirect.com/science/article/pii/S0360132325007887)
Abstract: Wind-driven rain (WDR) is one of the most primary moisture sources for building envelopes; however, existing semi-empirical models fail to rapidly and accurately calculate the WDR amount, particularly during rain events with intermittent periods and significant wind field fluctuations. This study explored the applicability and accuracy of machine learning methods in addressing these limits and compared their performance with the ASHRAE model. Five machine learning models were constructed to predict the WDR amount on the building facade with the field measurement dataset collected in Chongqing between 2022 and 2023. The results indicated that all machine learning models outperformed the ASHRAE model in predicting the WDR amount, with the most accurate machine learning prediction being 1.14 times the true value, significantly better than the ASHRAE model’s 1.98 times, and hyperparameter optimization further enhanced machine learning models’ performance. Specifically, the RBF neural network and the support vector machine demonstrated good generalization, stability and accuracy, whereas the random forest model exhibited instability issues. The BP neural network model performed the least effectively; however, it can be improved by refining its weights and thresholds through optimization algorithms. The study underscores the feasibility of machine learning methods and highlights the critical role of hyperparameter tuning in achieving optimal accuracy and reliability.
Keywords: Wind-driven rain; Machine learning model; Field measurement; ASHRAE model