基于动态权重系数多目标优化组合的贝叶斯渗流参数反演及分析  

Inversion and analysis of Bayesian seepage parameters based on multi-objective optimization combination of dynamic weight coefficients

作　　者：翟于广 任杰[1] 南胜豪 郭恒乐 陈凯旋 睢佳衡 ZHAI Yuguang;REN Jie;NAN Shenghao;GUO Hengle;CHEN Kaixuan;SUI Jiaheng(State Key Laboratory of Eco-hydraulics in Northwest Arid Region,Xi′an University of Technology,Xi′an,Shaanxi 710048,China)

机构地区：[1]西安理工大学省部共建西北旱区生态水利工程国家重点实验室,陕西西安710048

出　　处：《排灌机械工程学报》2024年第12期1259-1265,共7页Journal of Drainage and Irrigation Machinery Engineering

基　　金：国家自然科学基金资助项目(52379134);陕西省科技创新团队项目(2022TD-01)。

摘　　要：现有的贝叶斯参数反演方法普遍存在计算耗时过长、计算精度较低、计算准确性较差等问题,对此基于多次尝试差分进化自适应Metropolis(MT-DREAM(ZS))算法,构建了一种更加合理的确定单一模型权重系数组合代理模型,并通过基于Pareto最优的动态权重系数多目标优化组合对模型进行修正.在贝叶斯方法中通过集成多元自适应回归样条(MARS)、人工神经网络(ANN)、随机森林(RF)这3种机器学习方法构建组合模型,并在充分考虑反演过程中不确定性的基础上,对渗流参数的后验分布情况进行推导.结合实际工程的监测数据,通过计算预测性能指标决定系数(R^(2))和均方根误差(RMSE),对比分析该组合代理模型与其他模型的差距.结果表明,该组合代理模型的拟合精度高,预测效果好,相较于其他模型的提升幅度平均为15.00%~20.00%.将反演所得渗流参数运用到仿真模拟试验中的研究成果,为大坝渗流检测领域的发展提供了一种新的思路.Prevailing Bayesian parameter inversion techniques have often been marred by extended computational durations,diminished computational precision and sub-optimal accuracy.Hence,a hybrid surrogate model underpinned by the multiple attempts of differential evolution adaptive Metropolis(MT-DREAM(ZS))algorithm was introduced,which offered a more scientifically grounded approach for determining the weight coefficients of individual models,and was modified through Pareto optimization-based dynamic weight coefficient multi-objective optimization.Three distinct machine learning methodologies including multivariate adaptive regression splines,artificial neural network random forest,and random forest were integrated into the Bayesian framework to establish a composite model.Additionally,the posterior distribution of seepage parameters was deduced,while thoroughly accounting for uncertainties present in the inversion procedure.Combined with the monitoring data of the actual project,the gap between this combinatorial surrogate model and other models was compared and analyzed by calculating the prediction performance index R^(2) and RMSE.Research findings substantiate that the hybrid surrogate model,coupled with the novel technique for weight determination of individual models,boasts superior fitting precision and predictive efficacy.Compared with the traditional method,the improvement rate is 15.00%-20.00%on average.By applying the inverted seepage parameters to simulation experiments,a new approach is provided for the development of dam seepage detection research.

关 键 词：贝叶斯理论 参数反演 MT-DREAM_((ZS))算法 组合代理模型 多目标优化 

分 类 号：TV139.1[水利工程—水力学及河流动力学]