HYBRID PKO-XGBOOST MODEL FOR ACCURATE SHEAR STRENGTH PREDICTION OF CONCRETE-FILLED STEEL TUBE COLUMNS
Concrete-filled steel tubes (CFST) are extensively used in modern civil engineering structures because of their excellent load-bearing capacity, ductility, and seismic resistance. The composite interaction between the steel tube and the concrete core significantly enhances structural performance compared to conventional steel or reinforced concrete members. However, existing design standards, including the American Institute of Steel Construction (AISC) and Eurocode 4 provisions, often provide conservative shear strength predictions because they do not fully capture the complex composite mechanisms governing CFST behavior.
Limitations of Conventional Design Methods
Traditional empirical or semi-empirical design equations rely on simplified assumptions about material interaction and stress distribution. As a result, these models often underestimate the actual shear capacity of CFST columns, leading to conservative structural designs and inefficient material usage. Furthermore, such approaches cannot easily adapt to large experimental datasets or capture nonlinear relationships among structural parameters. This limitation motivates the adoption of data-driven predictive models capable of learning complex interactions within structural systems.
Hybrid Machine Learning Framework
To improve prediction accuracy, this study proposes a hybrid modeling approach that integrates Extreme Gradient Boosting (XGBoost) with the Pied Kingfisher Optimizer (PKO). PKO is a nature-inspired optimization algorithm designed to enhance model performance by efficiently tuning hyperparameters. By combining the strong learning capability of XGBoost with PKO's optimization strategy, the hybrid model achieves improved predictive capability and robustness when estimating the shear strength of CFST columns.
Prediction Interval and Error Mitigation Techniques
Beyond point predictions, the study incorporates quantile regression to generate prediction intervals for ultimate shear force, enabling uncertainty quantification in structural predictions. Additionally, the Asymmetric Squared Error Loss (ASEL) function is introduced to reduce the risk of overestimation errors, which are critical in structural safety evaluations. This approach ensures that the predictive model remains both accurate and conservative where necessary, aligning with engineering safety requirements.
Model Performance and Comparative Analysis
Computational results demonstrate that the PKO-XGBoost model significantly outperforms conventional models. The hybrid framework achieves a Mean Absolute Percentage Error (MAPE) of 4.431% and a coefficient of determination (R²) of 0.9925 on the test dataset. Furthermore, the ASEL-PKO-XGBoost variant effectively reduces overestimation errors to 28.26% while maintaining comparable predictive performance. These results confirm the effectiveness of the proposed framework for accurately predicting CFST shear capacity.
Development of Predictive Equation and Practical Tools
In addition to the machine learning model, a new strength equation was derived using a Genetic Algorithm (GA) combined with existing equation-based models. The resulting equation demonstrates improved accuracy (R² = 0.934) compared with traditional design formulas. To facilitate practical implementation, web-based Graphical User Interfaces (GUIs) were also developed, enabling engineers to perform real-time shear strength predictions efficiently. These tools support practical adoption of advanced predictive methods in structural design and engineering practice.
Comments
Post a Comment