Evaluating Machine Learning Models for Predicting Structural Subsidence Based on Leveling Data: A Case Study in Vietnam

Abstract

Monitoring and predicting structural subsidence is crucial for construction project safety and efficiency, particularly in regions like Vietnam where leveling-based monitoring is standard. This study evaluates four machine learning methods, Linear Regression (LR), Support Vector Machine (SVM), Random Forest (RF), and Gradient Boosting (GB) for predicting subsidence using leveling data from a high-rise building in Hanoi. The dataset comprised 11 measurement cycles from three monitoring points between July 2020 and September 2022. Performance metrics revealed RF as the most effective model, consistently yielding superior predictive accuracy. LR also demonstrated steady, practical performance. Conversely, GB and SVM performed poorly, likely constrained by the limited dataset size. Notably, RF showcased the potential to surpass traditional predictive approaches, offering a robust solution even with sparse data, while LR remains a viable option for resource-constrained scenarios. This research introduces a modern ML-based approach to subsidence prediction relevant to the Vietnamese context, highlighting the importance of dataset characteristics and underscoring the need for larger datasets and the inclusion of more influencing factors in future investigations to further refine predictive capabilities.