데이터 증강기법과 서포트 벡터 회귀를 활용한 초고층 건물 커튼월 공사 기간 예측

Abstract

Predicting the duration of high-rise building projects is essential for managing progress and identifying potential delays, especially for activities on the critical path that can disrupt schedules. Curtain wall installation is one of the key tasks that often cause delays. Although data-driven approaches show promise for accurate predictions, data scarcity in South Korea limits their application. To address this, exploring new data augmentation techniques and prediction methods is necessary. This study compares three Monte Carlo simulation (MCS) variants and the Synthetic Minority Over-Sampling Technique (SMOTE) for data augmentation, using data from 15 real projects. The augmented data is then analyzed with Support Vector Regression (SVR) using three different kernels. The model's accuracy is assessed using mean square error (MSE) and by comparing predicted durations with actual construction timelines. Results show that SMOTE combined with SVR linear yielded the lowest MSE at 0.047, while SMOTE with SVR radial basis function provided the most accurate prediction, with just a one-day error. These findings suggest that combining data augmentation techniques with machine learning can effectively address data limitations and improve forecasting of construction duration.