Comparative Analysis of Hyperparameter Optimization Techniques for PCA–XGBoost Models in SRCFSST Column Load Prediction

This study aims to predict the peak axial load capacity (PALC) of steel-reinforced concrete-filled square steel tubular (SRCFSST) columns under elevated temperatures using a machine learning-based approach. The motivation arises from the limitations of traditional experimental and numerical methods, which are often time-consuming, costly, and computationally intensive.

A hybrid predictive framework was developed by integrating Principal Component Analysis (PCA) with Extreme Gradient Boosting (XGB). The dataset, comprising 135 instances from prior experimental studies, underwent PCA for dimensionality reduction, retaining 99% of the variance. The PCA-transformed data was used to train XGB models, with hyperparameter tuning conducted via Grid Search, Random Search, and Bayesian Optimization. A 5-fold cross-validation technique was employed to enhance model generalizability, and performance was evaluated using R², RMSE, and WMAPE.

Among the three tuning strategies, the Bayesian-optimized PCA-XGB model demonstrated the highest predictive performance with an R² of 0.928, MAE of 2.3%, and RMSE of 3.5% on the test dataset. Statistical analyses, including paired t-tests and Wilcoxon signed-rank tests, confirmed the superiority of this model with significant improvements over other configurations (p < 0.05). The use of PCA notably reduced multicollinearity and computational complexity.

The findings underscore the value of combining dimensionality reduction with advanced hyperparameter tuning to develop efficient, accurate, and interpretable models for structural fire engineering. The PCA-XGB-BO framework offers a viable alternative to traditional modeling approaches, particularly for complex prediction problems involving high-temperature effects on structural components.

This study establishes a robust data-driven methodology for estimating PALC in SRCFSST columns exposed to high temperatures. The integration of PCA and Bayesian Optimization within an XGB modeling framework delivers high accuracy while reducing computational burden. Future research should focus on extending this framework to other structural systems, incorporating physics-informed constraints, and validating performance through large-scale experimental testing.