A hybrid semi-supervised regression based machine learning method for predicting peak wind loads on a group of buildings. (15th January 2023)

• Record Type: Journal Article
• Authors: Du, Kun; Chen, Bo
• Is Part Of: Engineering structures. Volume 275(2023)Part A
• Publication Date: 2023-01-15
• Subjects: Peak wind loads -- A group of buildings -- Semi-supervised learning -- Extreme learning machine -- Computational fluid dynamics; Structural engineering -- Periodicals; Structural analysis (Engineering) -- Periodicals; Construction, Technique de la -- Périodiques; Génie parasismique -- Périodiques; Pression du vent -- Périodiques; Earthquake engineering; Structural engineering; Wind-pressure; Periodicals; 624.105
• DOI: 10.1016/j.engstruct.2022.115245
• Languages: English
• ISSNs: 0141-0296

Abstract

Highlights: Proposes hybrid machine learning method SEC to assess peak wind loads on buildings. The SEC method integrates the SSR, ELM, and CFD methods. SEC uses SSR to expand limited wind tunnel samples to improve prediction precision. SEC has better prediction precision than ELM trained based on limited existing samples. Abstract: This study proposes a new hybrid machine learning method (named as SEC) to predict peak wind loads on a group of buildings. The SEC method integrates the semi-supervised regression, extreme learning machine, and computational fluid dynamics (CFD). Wind loads on a group of buildings are sensitive to many influencing factors, and there are not sufficient enough wind tunnel test samples for machine learning prediction. SEC uses semi-supervised regression to construct pseudo-labeled samples of untested buildings and mixes them with existing samples to expand training samples to improve the prediction precision. Wind tunnel tests demonstrate that peak wind loads sometimes have a strong relationship with the building spacing or alter rapidly with the variation of building spacings. CFD can calculate mean wind loads much more accurately than peak wind loads at low costs. SEC adds mean wind loads of untested buildings simulated by CFD as the input variable in the prediction model to weaken the nonlinear relationship between the input variables and peak wind loads and further improve the prediction accuracy. The results of the case study of a group of three flat-roof buildings show that the prediction performance of the SEC model using 2891 existing wind tunnel test samples reaches that of the extreme learning machine model using 5782 existing wind tunnel test samples when the predicted target is in the local range where peak wind loads alter rapidly with the increase of the building spacing. The prediction performance of the SEC model is improved with the increase in the number of pseudo-labeled samples.