Machine learning (ML) algorithms have been gradually used in predicting tunneling-induced settlement, but there is no uniform process for establishing ML models and even obviously exists deficiency in the existing settlement prediction ML models. This study systematically demonstrates the process of application of machine learning (ML) algorithms in predicting tunneling-induced settlement. The whole process can be categorized into four phases: the selection of ML algorithms, the determination of optimum-hyper-parameters, the improvement in model robustness and sensitivity analysis. The prediction performance of five commonly used ML algorithms back-propagation (BPNN), general regression neural network (GRNN), extreme learning machine (ELM), support vector machine (SVM) and random forest (RF) was comprehensively compared. The results indicate that proposed hybrid intelligent algorithm with the integration of the meta-heuristic algorithm particle swarm optimization (PSO) and ML can effectively determine the global optimum hyper-parameters of ML algorithms. The mean prediction error of k-fold cross-validation sets defined as the fitness function of the PSO algorithm can improve the robustness of ML models. RF algorithm outperforms the remaining four ML algorithms in recognizing the evolution of tunneling-induced settlement. BPNN shows great extrapolation capability, so it is recommended to establish settlement prediction model if the existing datasets are small. Sensitivity analysis indicates the geological and geometric parameters are the most influential variables for the settlement.

中文翻译：

---

用于隧道诱导沉降预测的混合元启发式和机器学习算法：比较研究

机器学习（ML）算法已逐渐应用于隧道沉降预测，但目前并没有统一的机器学习模型建立流程，甚至现有的沉降预测机器学习模型存在明显不足。本研究系统地展示了机器学习 (ML) 算法在预测隧道引起的沉降中的应用过程。整个过程可以分为四个阶段：ML算法的选择、最优超参数的确定、模型鲁棒性的提高和敏感性分析。综合比较了五种常用的ML算法反向传播（BPNN）、一般回归神经网络（GRNN）、极限学习机（ELM）、支持向量机（SVM）和随机森林（RF）的预测性能。结果表明，所提出的融合元启发式算法粒子群优化（PSO）和机器学习的混合智能算法可以有效地确定机器学习算法的全局最优超参数。定义为 PSO 算法适应度函数的 k 折交叉验证集的平均预测误差可以提高 ML 模型的鲁棒性。RF 算法在识别隧道诱导沉降的演变方面优于其余四种 ML 算法。BPNN 表现出很强的外推能力，因此如果现有数据集较小，建议建立沉降预测模型。敏感性分析表明地质和几何参数是对沉降影响最大的变量。