THE GATHERING OF A DATASET FOR TUNNELS AND THE ASSESSMENT OF PREDICTED CONSTRUCTION DELAY MODELS UTILIZING REGRESSION AND ADAPTIVE BOOSTING TECHNIQUES

Abstract

Construction delays in tunnel projects have persisted over several decades, often resulting in significant financial and scheduling impacts. Despite extensive efforts, the root causes of these delays and effective predictive modeling approaches remain insufficiently resolved. This study aims to identify the key factors contributing to construction delays and to develop predictive models based on empirical data from tunnel projects in Japan constructed using the New Austrian Tunneling Method (NATM). The dataset includes initial and final displacements, displacement rate, categorical geological classifications, and advance rate (dependent variable), compiled from detailed design and construction records. Descriptive statistical analysis revealed a high frequency of outliers and a non-normal distribution, suggesting underlying heterogeneity in ground conditions. Regression models—both standalone and integrated with K-means clustering—were developed and further refined using Adaptive Boosting (Adaboost) algorithms. Adaboost outperformed other models, achieving higher coefficients of determination (R²) and lower prediction errors. Feature importance and SHAP analysis confirmed final displacement as the most influential predictor of tunneling performance. The principal causes of delay were identified as insufficient geotechnical investigations and unanticipated disaster-related ground instabilities, both of which contributed to design revisions and prolonged construction periods. The study underscores the critical role of comprehensive geological surveys conducted at early project stages and demonstrates the utility of machine learning in enhancing delay prediction. These findings provide actionable insights for improving schedule reliability and risk management in future tunnel infrastructure development.