Tree-based machine learning models for prediction of bed elevation around bridge piers

Abstract

Scouring around bridge piers is a highly nonlinear process making its prediction by deterministic and stochastic models challenging. This study explores the application of inferential models for predictions of bed elevations around bridge piers. The objective is to get a generalized machine learning model with an interpretable structure. The historical data comprise a detailed record of streamflow and bed elevations that were captured by sensors installed at the 5th Street Bridge piers over Ocmulgee River at Macon, GA. We investigate the accuracy and efficiency of various tree-based machine learning algorithms, including a single tree as well as homogeneous ensemble models for simultaneous predictions of bed elevation at multiple sensors installed at piers. The ensemble models were based on bagging and boosting techniques. Special attention is given to balancing between overfitting and underfitting without compromise on the model's robustness. Observation of the performance metrics showed that tree-based models have excellent predictive capacity. It was observed that boosting models, including a gradient based regression model, and adaptive boosting outperformed the bagging model. Among all the models investigated in this study, the adaptive boosting method was observed to be most generalizable. The performance of developed models shows the potential of tree-based ensemble models in providing rapid and robust predictions for complex nonlinear fluid flows. Published under an exclusive license by AIP Publishing.