An Intelligent Prediction Model for Strain of Station Pipelines Under Settlement Based on Finite Element Simulation and Stacking Ensemble Learning

Abstract

Settlement-induced bending may cause excessive tensile and compressive strains in buried station pipelines, while full finite element analysis is too time-consuming for rapid integrity screening. This study proposes a strain prediction framework that couples nonlinear pipe–soil finite element simulation with stacking ensemble learning. A pipe–soil model for X65 buried pipelines is established to generate 196 samples with pipe diameter, wall thickness, settlement length, and settlement amount as inputs, and maximum tensile and compressive strains as outputs. Random Forest, LightGBM, and support vector regression are trained as base learners and then fused through stacking. Results show that RF performs best for tensile strain prediction (R2 = 0.8815), whereas SVR performs best for compressive strain prediction (R2 = 0.8997). The stacking models further improve accuracy, with RF + LGBM + SVR achieving R2 = 0.9033 for tensile strain and LGBM + SVR achieving R2 = 0.9048 for compressive strain. The proposed model provides an efficient tool for settlement pipeline integrity assessment.