Intelligent Model for the Compressibility Prediction of Cement-Stabilized Dredged Soil Based on Random Forest Regression Algorithm

Abstract

The determination of the compression characteristics of cement-stabilized dredged soil is an important issue. However, conventional method of laboratory compression test is time-consuming. Considering its robustness, random forest (RF) regression algorithm was utilized in this research to predict the compressibility of cement-stabilized dredge soil. The input variables included fundamental physical parameters of uncemented soil, water/cement ratio, and cement content. The output variables were initial void ratio e0 and compression index Cs at the pre-yield state, which can be used to describe the compression curve. Furthermore, a comparative analysis between the RF and support vector machine (SVM) algorithms was implemented, and variable importance was also evaluated according to the RF model. Results indicate that the RF model could quickly predict Cs and e0 values of cement-stabilized dredged soil without one-dimensional oedometer tests. The RF algorithm was found to exhibit better performance than the SVM algorithm, and it has a strong ability to avoid over-fitting. The determination coefficient R2 for Cs and e0 on the validation set were found to respectively reach 0.83 and 0.97. Moreover, the importance values of the initial water content, water/cement ratio, and plastic limit were found to be 0.313, 0.249, and 0.117, respectively. The secant compression modulus Es1−2 obtained from the compression curves could be applied to the calculation of settlement and stress in engineering design.