【摘要】 在流变软岩中进行水工隧道施工时，围岩-衬砌相互作用将在接触面处产生与时间相关的支护力。合理预测支护力对衬砌结构设计和围岩稳定性分析至关重要。然而，黏弹塑性问题的数值模拟需要长时间的计算迭代和复杂的前后处理，而解析方法则面临公式推导与程序实现的难题，限制了推广应用。基于衬砌支护隧道的黏弹塑性解析模型生成大量数据集，将围岩和衬砌材料特性及几何参数等6个核心参数作为特征，利用LightGBM机器学习方法建立了稳态支护力的数据驱动预测模型。结果表明：该模型在测试集上的预测效果稳定，训练集与测试集的校正决定系数均超过0.9,平均绝对百分比误差低于4.2%,显著优于XGBoost及SVR等其他机器学习算法。此外，利用SHAP方法分析了输入特征与预测结果之间的依赖性，增强了模型的可解释性。该预测模型不仅能快速准确地进行稳态支护力预测，还能应用于衬砌设计和其他反分析工作，具有较高的工程应用价值。更多还原

【Abstract】 For hydraulic tunnelling in rheological soft rocks, time-dependent support forces are likely to occur at the rock-support interface due to the interactions between the surrounding rocks and the support structures. A better understanding of the mechanism is crucial for the design of lining structures and the stability analysis of surrounding rocks. Numerical simulations of viscoelastic-plastic problems are time-consuming due to the large consumption of computational iterations and pre-/post-processing; meanwhile, the analytical approach needs complex formula derivation and program implementation. Therefore, both methods are far from satisfaction when it comes to engineering applications. In this study, a big group of datasets are obtained based on the viscoelastic-plastic analytical model for supported tunnels, with six different key parameters of material properties and geometry information of the surrounding rocks and support structures as the feature parameters. Subsequently, a data-driven model for predicting steady-state support forces is developed by LightGBM machine learning method. The results indicate that the model demonstrates stable predictive performance on the test set, with calibration determination coefficients for both the training and test sets exceeding 0.9, and mean absolute percentage errors below 4.2%, significantly outperforming other machine learning algorithms such as XGBoost and SVR. Finally, SHAP analysis is applied to assess the relationship between input features and predicted results, further enhancing the model’s interpretability and providing deeper insights into feature contributions. In summary, the developed data-driven model can rapidly predict steady-state support forces, and it can be further used in the design of support structures and other works of inverse analyses.更多还原