Monitoring and predicting the displacement of concrete dams is one of the most crucial considerations for ensuring their long-term safe operation. Most existing models are designed for dams subjected to common structural and environmental conditions, with little attention paid to atypical operational conditions, such as structural strengthening or sudden changes in the external environment. The motivation for this work is to develop a reliable prediction model for displacement behavior of concrete dams subjected to irregular upstream water-level fluctuations. In our study, a reconstruction method for unevenly sampled time series is presented to analyze such data. Then, an improved model, factor weighted support vector regression (FWSVR), which differentiates various factors and their effects through a weighting matrix, is theoretically derived. The weights are determined by the RReliefF algorithm, and together with FWSVR form the final RReliefF-based FWSVR (RFWSVR) model. In particular, the hyperparameters involved in the above modeling strategy are optimized by the Grey Wolf Optimizer. Eventually, the prediction robustness of the developed model was verified on the data from four representative monitoring points of a real-world dam, where its accuracy was compared to classical dam behavior modeling methods, FWSVR models using other weighting methods, and an ensemble learning algorithm. Comparative evaluation of the performance of the different methods was conducted with the help of recognized statistical indices. The evaluation results show that the overall performance of the proposed RFWSVR model is optimal for the displacement prediction at the selected points when the dam case is subjected to irregular water-level changes. This novel modeling approach may be generalized for modeling the evolution behavior of other civil or hydraulic structures.

中文翻译：

---

水位不规则波动下混凝土大坝位移的鲁棒预测模型

监测和预测混凝土大坝的位移是确保其长期安全运行的最关键考虑因素之一。现有的大多数模型都是为受普通结构和环境条件影响的大坝设计的，很少关注非典型操作条件，例如结构加固或外部环境的突然变化。开展这项工作的动机是为受不规则上游水位波动影响的混凝土大坝的位移行为建立可靠的预测模型。在我们的研究中，提出了一种用于采样时间序列不均匀的重构方法来分析此类数据。然后，理论上得出了一种改进的模型，即因子加权支持向量回归（FWSVR），该模型通过加权矩阵来区分各种因子及其影响。权重由RReliefF算法确定，并与FWSVR一起形成最终的基于RReliefF的FWSVR（RFWSVR）模型。特别是，上述建模策略中涉及的超参数是由Gray Wolf Optimizer优化的。最终，在真实大坝的四个代表性监测点的数据上验证了开发模型的预测鲁棒性，并与经典大坝行为建模方法，使用其他加权方法的FWSVR模型和集成学习算法进行了比较。 。在公认的统计指标的帮助下，对不同方法的性能进行了比较评估。评价结果表明，当大坝工况受到不规则水位变化时，所提出的RFWSVR模型的整体性能对于选定点的位移预测是最佳的。可以推广这种新颖的建模方法，以对其他土木或水工结构的演化行为进行建模。