Performance Comparison of Conventional Machine Learning and Stacking Ensemble Models for Predicting Peak Seismic Displacement Response of Columns Considering Geometric Parameters

형상 변수에 따른 기둥의 최대 지진 변위응답 예측을 위한 기존 머신러닝과 스태킹 앙상블 모델의 성능 비교

This study comparatively analyzes conventional machine learning models and a stacking ensemble model for predicting the peak displacement response of concrete columns under seismic loading. Using geometric parameters, material properties, and ground motion characteristics as inputs, a database of 1,080 cases was constructed via finite-element time-history analysis. Hyperparameter optimization was applied to six conventional machine learning models based on tree-based, kernel-based, and neural network approaches. A stacking ensemble is constructed with Extra Trees and XGBoost as base models and RidgeCV as the meta-model, and its optimal configuration is derived by examining meta-model selection, passthrough application, and the number of cross-validation folds. Predicted peak displacement responses were converted into drift ratios normalized by column height to evaluate predictive performance across models. The tree-based models show stable performance with a MAPE of 3-4%, while the stacking ensemble achieves the lowest error rate at 2.68% MAPE. Interval-based analysis shows that the stacking ensemble yields lower error rates than conventional machine learning models across a range of geometric parameters, including column height, cross-sectional dimensions, and aspect ratio, confirming stable predictive performance under variations in these parameters.