Performance comparison of machine learning algorithms for maximum displacement prediction in soldier pile wall excavation

Abstract

One of the common excavation methods in the construction of urban infrastructures as well as water and wastewater facilities is the excavation through soldier pile walls. The maximum lateral displacement of pile wall is one of the important variables in controlling the stability of the excavation and its adjacent structures. Nowadays, the application of machine learning methods is widely used in engineering sciences due to its low cost and high speed of calculation. This paper utilized three intelligent machine learning algorithms based on the excavation method through soldier pile walls, namely eXtreme gradient boosting (XGBoost), least square support vector regressor (LS-SVR), and random forest (RF), to predict maximum lateral displacement of pile walls. The results showed that the implemented XGBoost model performed excellently and could make predictions for maximum lateral displacement of pile walls with the mean absolute error of 0.1669, the highest coefficient of determination 0.9991, and the lowest root mean square error 0.3544. Although the LS-SVR, and RF models were less accurate than the XGBoost model, they provided good prediction results of maximum lateral displacement of pile walls for numerical outcomes. Furthermore, a sensitivity analysis was performed to determine the most effective parameters in the XGBoost model. This analysis showed that soil elastic modulus and excavation height had a strong influence on of maximum lateral displacement of pile wall prediction.