地下水模拟不确定性问题的多模型分析

Multiple Model Analysis for Studying Groundwater Uncertainties

为研究地下水概念模型的构建偏差及水文地质参数非均质性引起的地下水渗流场模拟不确定性问题,首先根据自然条件的差异构建2组概念模型;以大量原位水文地质试验获取的待估参数数据为先验信息,应用接受条件进行调整的马尔科夫链蒙特卡罗方法(MCMC)中的自适应采样算法(A-M)进行参数样本采集,并基于2组概念模型分别构建多组渗流场计算模型;将输出结果基于AICc准则进行相关多模型定量分析.研究结果表明:调整的A-M采样算法,参数样本的遍历性及收敛性未受影响;计算模型中除存在"异参同效",亦存在"异构同效";异构同效虽存在,但更接近客观条件的概念模型结构获取高精度模型的概率较大,1#、2#概念模型中方差值介于1~2的模型比例分别为65%、46%;各概念模型的100组计算模型中,剔除Delta值大于10的计算模型后,1#模型中仅保留排名前10个模型,累计后验概率0.996,2#模型则保留排名前21个模型,而其排名前10的模型累计后验概率仅为0.884.

Multiple Model Analysis was applied to study the groundwater modelling uncertainties caused by the deviation of model structure and heterogeneity in aquifer media. According to different natural conditions,two hydrogeological conceptual models were established. Using a large number of model parameter data,obtained through hydrogeological tests,as a priori information and based on the two conceptual models,a series of seepage field models was constructed using the Adaptive MetropolisMarkov Chain Monte Carlo method that acceptance condition was adjusted. Uncertainties of modelling output data are analysed based on corrected Akaike＇s Information Criteron. Research indicates that the ergodicity and convergence of sample parameters will not be affected by changes in acceptance conditions. The model output data include the following effects： ＂same results with different parameters＂and ＂same results with different models ＂. Although these effects exist, the model structure is closer to the objective of improving the probability of obtaining a high precision model. The proportion of the primary conceptual model,with a variance between 1 and 2,is 65%. When the model with Delta values greater than 10 is excluded,the top 10 models are retained and the cumulative a posterior probability is 0. 996. The proportion of the second conceptual model,with a variancebetween 1 and 2,is 46%. When the model with Delta values greater than 10 is excluded,the top 21 models are retained. The cumulative posterior probability of the top 10 models is only 0. 884.