基于多因子融合和Stacking集成学习的大坝变形组合预测模型

变形是反映大坝服役形态变化的直观表征，构建高效准确的变形预测模型对于大坝结构安全控制十分重要。传统单因子及单算法变形预测模型存在泛化能力不足、鲁棒性差等问题，易出现预测偏差甚至误判。针对这一问题，本文选取不同变形解释因子及回归算法，构建多种单因子单算法预测模型，结合Stacking集成学习思想，对上述模型进行组合，提出了大坝变形组合预测模型。该组合模型以Stacking集成学习为核心，采用高斯过程回归作为元学习器，从算法、因子两方面对单因子单算法预测模型进行集成，并通过k折交叉验证减小模型过拟合风险。以某混凝土拱坝变形监测数据为例，通过多模型构建与性能比较，对所提出模型的准确性与有效性进行评估。结果表明：单因子单算法预测模型具备准确性和多样性的特征；通过算法、因子集成，组合模型的预测精度和鲁棒性得到了显著提高，在水位波动期的预测能力得到了有效增强。综上，大坝变形组合预测模型具备出色的非线性信息挖掘与建模预测能力，可为大坝安全监测提供可靠依据。

Combined prediction model of dam deformation based on multi-factor fusion and Stacking ensemble learning

Deformation is the intuitive reflection of the change of dams'' operating behavior.It''s crucial to build a deformation prediction model with higher efficiency and accuracy for dam structural safety monitoring.Traditional single-factor and single-algorithm prediction models inevitably have series of problems such as insufficient generalization ability and poor robustness, which will induce deviations and even misjudgments.To solve this problem, this paper selects different deformation interpretation factors and regression algorithms to build multiple single-factor single-algorithm prediction models.Next, these models are integrated to propose a combined dam deformation prediction model through Stacking ensemble learning.Detailedly, this combined model adopts Gaussian Process Regression as the meta-learner and integrates the single-factor single-algorithm models from algorithm and factor these two aspects.To reduce the risk of overfitting, k-fold cross-validation is also introduced in generating the new data set.Referring to the deformation data of a concrete arch dam, the model''s accuracy and effectiveness have been evaluated by multi-model construction and performance comparison.The results show that the single-factor single-algorithm models are characterized by accuracy and diversity.Through the integration of algorithms and factors, the prediction accuracy and robustness have been significantly improved, and the prediction capability of the models has been effectively enhanced during the water-fluctuating period.Above all, the combined dam deformation prediction model has excellent nonlinear information mining ability and predictive performance, and could provide a reliable basis for dam safety monitoring.