The purpose of this study is to compare the prediction performances of different commonly used data mining models for high arch dam deformation prediction during the initial operation period. Four machine learning modeling techniques, namely random forest (RF), least squares support vector machines (LS-SVM), simple boosted regression tree (SBRT), kernel extreme learning machine (K-ELM), and modified statistical models are used. With the help of hierarchical clustering on principal component analysis, the typically measured temperatures representing the unsteady dam temperature field are selected to reduce dimensionality, avoid over-fit, and facilitate explanation. The practical term for the time-dependent effect, reflecting both the exponential growth and the recovery term, is also introduced. These models are constructed, and their prediction and explanatory capabilities to model dam deformation are compared by dividing time series into training and test sets. Four different training and test combinations are investigated. The traditional hydrostatic–seasonal–time (HST) model is not enough to predict this nonlinear deformation. The SBRT, the modified HST, and its improved HT_TTS (hydrostatic–thermal by measured temperature–time–season model) models show good performance. Relative influence and partial dependence plot are employed to understand the relationship between the deformation and the predictor variables. The results show that several factors, including reservoir thermal stratification, concrete temperature rise, valley contraction, and non-monotonic time-dependent effect, affect the deformation of high arch dams during the initial operation period.

这项研究的目的是为了比较不同的常用数据挖掘模型在初始运营期间对高拱坝变形预测的预测性能。使用了四种机器学习建模技术，即随机森林（RF），最小二乘支持向量机（LS-SVM），简单增强回归树（SBRT），内核极限学习机（K-ELM）和修改后的统计模型。借助基于主成分分析的层次聚类，可以选择代表不稳定大坝温度场的典型测量温度，以减小尺寸，避免过度拟合并简化说明。还介绍了时间相关效应的实用术语，它既反映了指数增长又反映了恢复期。这些模型被构建，通过将时间序列分为训练集和测试集，比较了它们对大坝变形建模的预测和解释能力。研究了四种不同的训练和测试组合。传统的静水-季节-时间（HST）模型不足以预测这种非线性变形。SBRT，改进的HST及其改进的HT_T TS（通过测得的温度-时间-季节模型得出的静水压力-热量）模型显示出良好的性能。使用相对影响力和偏相关图来了解变形与预测变量之间的关系。结果表明，包括水库热分层，混凝土温度升高，谷底收缩和非单调时效效应在内的几个因素会影响高拱坝在初始运营阶段的变形。