Scholarlink

<p>Predicting the axial load capacity of CFST columns is essential for structural and seismic safety. It enables the assessment of a building’s earthquake resistance and the establishment of post-earthquake safety standards. Axial load variations significantly affect the strength of RC columns, especially those made with low-strength concrete. This study employs advanced machine learning models, such as the Adaptive Neuro-Fuzzy Inference System (ANFIS) and Light Gradient Boosting Regression (LGBR), used individually or in combination to enhance prediction accuracy. Model parameters were optimized using the Hippopotamus Optimization Algorithm (HOA) and Dung Beetle Optimizer (DBO). The best-performing models were then combined into an ensemble system via Dempster–Shafer Theory (DST). The model was assessed via five-fold cross-validation and statistical indices (R<sup>2</sup>, MARE, RMSE, SMAPE, U95). A global sensitivity analysis method became necessary to find the main elements which impact the axial load capacity. The LGHO model (LGBR optimized with HOA) was the most stable, with a standard deviation of 17.638, about 54.8% lower than the baseline LGBR (39.079). FAST sensitivity analysis shows thickness (t) has the highest first-order effect (S1 = 0.06766), indicating the strongest contribution, while concrete compressive strength (f′c) has the dominant total effect (ST = 0.90583), confirming its critical influence on axial load capacity. The Wilcoxon signed-rank test found no significant performance differences among individual and hybrid models at 5% level (<i>p</i> &gt; 0.05), suggesting comparable approaches despite minor error variations. These results indicate that ensemble and optimization-enhanced AI models can improve axial load capacity prediction reliability in structural engineering.</p> Graphical abstract <p></p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Improved prediction of axial load capacity of CFST columns using machine learning algorithms and ensemble methods

  • 期刊论文

  • 发表日期：

    2026年3月16日

查看全文
  • Pengyu Zhang,
  • Meng Jia,
  • Xiaojuan Wei

摘要

Predicting the axial load capacity of CFST columns is essential for structural and seismic safety. It enables the assessment of a building’s earthquake resistance and the establishment of post-earthquake safety standards. Axial load variations significantly affect the strength of RC columns, especially those made with low-strength concrete. This study employs advanced machine learning models, such as the Adaptive Neuro-Fuzzy Inference System (ANFIS) and Light Gradient Boosting Regression (LGBR), used individually or in combination to enhance prediction accuracy. Model parameters were optimized using the Hippopotamus Optimization Algorithm (HOA) and Dung Beetle Optimizer (DBO). The best-performing models were then combined into an ensemble system via Dempster–Shafer Theory (DST). The model was assessed via five-fold cross-validation and statistical indices (R2, MARE, RMSE, SMAPE, U95). A global sensitivity analysis method became necessary to find the main elements which impact the axial load capacity. The LGHO model (LGBR optimized with HOA) was the most stable, with a standard deviation of 17.638, about 54.8% lower than the baseline LGBR (39.079). FAST sensitivity analysis shows thickness (t) has the highest first-order effect (S1 = 0.06766), indicating the strongest contribution, while concrete compressive strength (f′c) has the dominant total effect (ST = 0.90583), confirming its critical influence on axial load capacity. The Wilcoxon signed-rank test found no significant performance differences among individual and hybrid models at 5% level (p > 0.05), suggesting comparable approaches despite minor error variations. These results indicate that ensemble and optimization-enhanced AI models can improve axial load capacity prediction reliability in structural engineering.

Graphical abstract