| 摘要: |
| 为了提高大坝变形预测的精度与效率，针对大坝监测中存在的复杂非线性影响因素、时序动态特征利用不足及单一模型处理能力存在不足等问题，提出一种基于徒步优化算法（HOA）、时间卷积网络（TCN）、长短期记忆网络（LSTM）和极端梯度提升（XGBoost）的大坝变形预测模型。该模型构建了“深层特征提取—集成决策回归”的组合模型处理机制。首先利用TCN通过膨胀因果卷积提取局部时序特征，然后输入到LSTM捕捉长期时序演化规律；最后引入XGBoost将提取的深层时序特征与原始特征融合输出，提高模型精度。同时考虑组合模型参数难以选取，融合HOA优化算法对混合模型的超参数进行自适应寻优，避免人工调参的盲目性。以某双曲拱坝为例，构建多个组合模型，结果表明：该模型的预测性能明显优于其他组合模型，在不同测点上的相关系数R2分别达到0.998、0.992、0.997、0.994并具有更低的MAE、RMSE。综上，HOA-TCN-LSTM-XGBoost模型不仅提高了对复杂时变特征的解析能力，更有效克服了单一模型在特征利用率和预测稳定性上的不足，可以为大坝变形预测提供一个新方向。 |
| 关键词: 大坝变形预测 时间卷积网络 长短期记忆网络 极端梯度提升 徒步优化算法 |
| DOI： |
| 分类号: |
| 基金项目:国家自然科学基金项目（面上项目，重点项目，重大项目） |
|
| Dam deformation prediction model based on HOA-TCN-LSTM-XGBoost |
|
Zhang Haozhe, Zheng Dongjian, Liu Xiaoshuang, Cai An
|
|
Hohai University College of water conservancy and hydropower engineering
|
| Abstract: |
| The concrete dam is under the coupling effect of complex environment for a long time, and continues to bear the joint influence of key environmental factors such as water load and temperature cycle. Under this action, the dam structure will inevitably produce time-dependent deformation. This deformation process directly threatens the long-term operation safety of the dam and may weaken the economic benefits of the project. In order to solve these problems, this paper comprehensively analyzes the shortcomings of the existing research, and puts forward three innovative advantages: complementary time series modeling, hybrid architecture innovation and optimization algorithm call. The convolutional neural network (TCN), long-term and short-term memory network (LSTM) and extreme gradient lifting (XGBoost) are combined to extract the deep temporal features of dam deformation data, and the extracted deep temporal features are fused with the original data features to improve the prediction accuracy. In the aspect of super parameter selection, the global optimization and high-precision advantages of the walking optimization algorithm (HOA) are used to realize the super parameter optimization. The HOA-TCN-LSTM-XGBoost combination model was established.
Firstly, aiming at the abrupt change of the water level component in the impact factor, this model uses the expansion convolution and causal convolution structure of the TCN module to capture the multi-scale transient response characteristics, and relies on its powerful receptive field to capture the local abrupt change of the water level component in the dam deformation monitoring data; Then considering that TCN is not good enough in dealing with the long lag effect of temperature load, considering the periodic smooth evolution characteristics of dam deformation caused by periodic temperature load changes, the global trend between time series data is mined by using LSTM gating structure; Finally, the XGBoost model is constructed by fusing the deep temporal features extracted by the above two modules with the features of the original data, and the feature fusion output is carried out relying on its strong regularization constraint and efficient nonlinear mapping ability. At the same time, the number of convolution cores, the size of convolution cores, the number of residual blocks and the discard rate of TCN; The maximum number of iterations, the maximum depth of the tree of XGBoost have a great impact on the prediction performance of the model, the HOA optimization algorithm is used to optimize the combination of the model''s hyper parameters. The HOA algorithm takes the search space of the optimization algorithm as the terrain steepness and introduces the concepts of local peaks and global peaks as the definitions of local and global optima. The algorithm can achieve the balance between local optimization and global search, and has high adaptability under the benchmark function test.
In order to verify the superiority of HOA-TCN-LSTM-XGBoost combined model in dam deformation prediction, comprehensive comparative tests were carried out on four different measuring points of a concrete double curvature arch dam. The comparison test includes multiple benchmark models, including the single module model of TCN, LSTM, XGBoost and the current mainstream time series prediction models Transformer and Transformer-LSTM. The test establish multiple statistical indicators, such as mean square error (MSE), root mean square error (RMSE), mean absolute error (MAE) and determination coefficient R2 for objective evaluation. The results showed that the correlation coefficient R2 at different measuring points reached 0.998, 0.992, 0.997 and 0.994 respectively, and had lower MAE and RMSE. At the same time, in order to further evaluate the contribution of each component to the prediction effect of the combined model, ablation experiments were designed for this combined model. Ablation experiments verify the complementarity and necessity of the architecture of the combined model. To sum up, the test results show that the combined model combines the multi-scale feature extraction of TCN, the global temporal logic analysis of LSTM, the regularization constraint and the efficient nonlinear mapping ability of XGBoost to achieve the fusion output of deep temporal features and original features, and can use the advantages of each model to solve the problems of high nonlinearity and complex temporal dynamic characteristics of dam deformation data, and still has certain advantages compared with the mainstream temporal models, providing a new idea and direction for dam safety monitoring. |
| Key words: Dam deformation prediction Temporal Convolutional Network Long Short-Term Memory eXtreme Gradient Boosting Hiking Optimization Algorithm |
