Data-and experience-driven prediction method of twin shield tunneling-induced complicated settlement curve and its evolution

The application of machine learning algorithms in geotechnical engineering is increasingly prevalent. Solely data-driven machine learning algorithms suffer from the “black box” issue, lacking the ability to uncover causal relationships and exhibiting preferences in variable selection. Consequently, data-and experience-driven machine learning algorithms are gradually emerging. Addressing the prediction of complex ground settlement curves and their evolution, this paper proposes a novel data-and experience-driven machine learning prediction method. Specifically, it replaces the traditional “prediction of settlement points—settlement curve” approach with “prediction formula—settlement curve”. This method first selects an appropriate formula based on the characteristics of ground settlement curves, ensuring that formula parameters possess physical meanings. Subsequently, machine learning is employed to predict the parameters within the formula, which are then applied to derive the ground settlement curve. Sensitivity analysis of these parameters explores causal relationships and identifies model preferences. Finally, this approach is applied to predict ground settlement curves induced by twin shield tunnels excavation, using overlaid Peck curves as the prediction formula. GRNN and LSTM algorithms are employed to predict undetermined parameters within the ground settlement curve, yielding promising results that ensure accuracy in predicting key indices like maximum settlement while effectively capturing the overall shape of the settlement curve. This method enhances the explainability of machine learning in predicting ground settlement and provides valuable insights for forecasting the overall forms of complex ground settlement curves.