Prediction of soft soil compression performance based on improved stacking model

Abstract

[Objective] To establish a model for the rapid prediction of the compression coefficient, multiple easily obtainable physical property indicators are used as input features, and an improved stacking model based on principal component analysis(PCA) for dimensionality reduction is proposed. [Methods] Based on 360 sets of actual silty soil data collected from the Binhaiwan New District in Dongguan City, the model was trained and validated. To address the issues of high dimensionality of soil data and potential multicollinearity among features, PCA was first applied to extract the top five principal components with a cumulative variance contribution rate exceeding 95% from the feature data. These top five principal components mainly reflected the comprehensive variation characteristics of physical properties such as water content, natural void ratio, liquid limit, plasticity index, and specific gravity. The constructed stacking model consisted of two learning layers: base learners and a meta-learner. Random forest and XGBoost models were adopted as base learners, while a support vector regression(SVR) model was employed as the meta-learner. Out-of-fold predictions generated via cross-validation were used as input features for the meta-learner layer. [Results] The validation result indicated that the constructed stacking model achieved a coefficient of determination(R²) of 0.790 and a mean squared error(MSE) of 0.037 MPa^-2 on the training set, and an R² of 0.781 and an MSE of 0.033 MPa^-2 on the test set. Compared to traditional machine learning models, the proposed improved stacking model achieved higher prediction accuracy. Specifically, compared to the relatively better-performing random forest model, the R² increased by approximately 11.89%, while the MSE decreased by about 25%. Furthermore, a comparison was made with the model without PCA dimensionality reduction, which achieved an R² of 0.747 and an MSE of 0.048 MPa^-2 on the test set, both worse than the model using PCA for dimensionality reduction. [Conclusion] To enhance the model's generalization under varying data distributions across regions, a calibration method based on median mapping for prediction result is introduced. By comparing the median and variance relationship of the compression coefficient distributions between new and source regions, linear correction is applied to the prediction result, mitigating the impact of distribution drift on the model's prediction accuracy. The result demonstrate that the proposed model exhibits high reliability, generalizability, and application value in the rapid and accurate prediction of the compression coefficient in soft soil.

Keywords

compression coefficient of soft soil / principal component analysis(PCA) / stacking model / support vector regression(SVR) / random forest / XGBoost / cross-validation / mechanical properties