| 余红玲,王晓玲,王成,曾拓程,余佳,盖世聪.贝叶斯框架下大坝渗流参数反演组合代理模型研究[J].水利学报,2022,53(3):306-315,324 |
| 贝叶斯框架下大坝渗流参数反演组合代理模型研究 |
| Research on ensemble surrogate models of dam seepage parameters inversion under Bayesian framework |
| 投稿时间：2021-07-01 |
| DOI：10.13243/j.cnki.slxb.20210593 |
| 中文关键词: 渗流参数 贝叶斯反演 组合代理模型 DREAMZS算法 SVR Kriging MARS |
| 英文关键词: seepage parameter Bayesian inversion ensemble surrogate model DREAMZS algorithm SVR Kriging MARS |
| 基金项目:国家自然科学基金雅砻江联合基金项目(U1865204) |
|
| 摘要点击次数: 2257 |
| 全文下载次数: 918 |
| 中文摘要: |
| 渗流参数贝叶斯反演的关键在于解决对渗流正演模型大量调用而导致的计算耗时问题。现有提高贝叶斯反演计算效率的研究大多采用基于单一机器学习算法的代理模型，计算精度较低。针对上述问题，本文提出一种贝叶斯框架下大坝渗流参数反演组合代理模型。该方法在贝叶斯框架下集成支持向量回归（SVR）、Kriging和多元自适应回归样条（MARS）三种机器学习算法。其中，利用差分进化自适应Metropolis（DREAMZS）算法并行采样的优势计算权重系数的随机分布函数，在考虑不确定性的条件下获得模型权重系数。案例分析表明，相比于运行一次至少需要耗费4h的渗流数值模型，本文所提组合代理模型运行一次仅需几秒钟，显著提高了贝叶斯反演的计算效率；此外，本文所提反演方法相比于基于SVR、Kriging和MARS的贝叶斯反演方法能够获得更准确的反演结果，其平均精度分别提高了13.78%、19.34%和12.27%，为大坝渗流参数反演提供了一种新思路。 |
| 英文摘要: |
| The key to Bayesian inversion of seepage parameters is to solve the problem of time-consuming calculations caused by the large number of calls to seepage forward models. Most of the existing researches on improving the computational efficiency of Bayesian inversion adopt surrogate models based on a single machine learning algorithm，which has the problem of low computational accuracy. In response to the above problems，an ensemble surrogate model is proposed for the inversion of dam seepage parameters under the Bayesian framework in this paper. This method integrates Support Vector Regression （SVR）， Kriging and Multivariate Adaptive Regression Splines （MARS） machine learning algorithms under the Bayesian framework. Among them， the advantage of parallel sampling in the Differential Evolution Adaptive Metropolis （DREAMZS） algorithm is used to calculate the random distribution function of the weight coefficients， and the model weight coefficients are obtained under the consideration of uncertainty. Case analysis shows that compared to the seepage numerical model that takes at least 4 hours to run once，the ensemble surrogate model proposed in this paper takes only a few seconds，which significantly improves the computational efficiency of Bayesian inversion. In addition，compared with Bayesian inversion methods based on SVR，Kriging， and MARS， the Bayesian inversion method based on the ensemble surrogate model proposed in this paper can obtain more accurate inversion results，and its average accuracy has been increased by 13.78%， 19.34%，and 12.27% respectively，which provides a new idea for the inversion of dam seepage parameters. |
|
查看全文
查看/发表评论 下载PDF阅读器 |
| 关闭 |