【作者】 辛欣；

【导师】 卢文喜；

【作者基本信息】 吉林大学 ， 水文学及水资源， 2011， 博士

【摘要】 在石油的开采、炼制、储运和使用的过程中,由于泄漏、偷排和意外事故等原因,使原油和各种石油类产品进入环境而造成污染。石油产品主要是由烷烃、环烷烃和芳香烃组成的混合物,具有致癌、致畸和致突变的潜在威胁,属有毒污染物,对人类和环境都具有巨大的危害。石油类污染物在水中的溶解度一般很小,进入地下环境后通常以非水相流体(NAPLs,Non-Aqueous Phase Liquids)的形式存在。重非水相流体(DNAPLs)具有高密度、低水溶性和高界面张力的特性,比轻非水相流体(LNAPLs)更难修复,像常用的抽出-处理技术对它都难以奏效,并且修复费用非常昂贵,单个污染场地的去除修复费用常常需要数亿美元。近年来出现的表面活性剂冲洗技术,也称为表面活性剂强化含水层修复技术(Surfactant Enhanced Aquifer Remediation, SEAR),是对抽出-处理技术的改进。表面活性剂对憎水性有机污染物具有增溶作用(Solubilization)和增流作用(Mobilization),能有效提高DNAPLs在水中的溶解性和迁移性,能使更多的自由相的DNAPLs进入水中,从而大幅度提高抽出-处理技术对于DNAPLs修复的有效性。目前,表面活性剂强化含水层修复技术(SEAR)尚处在发展阶段,影响SEAR修复效果和修复费用的因素非常复杂,包括抽、注水井的选位,抽、注水强度的大小与分配,表面活性剂的浓度等。因此,如何在现场调查的基础上,通过模拟模型和优化模型的合理运用,对修复工程方案进行优选,以提高修复效率并节省修复费用,是一个亟待解决且具有重要理论和实际意义的科学问题。而在运用模拟模型和优化模型的过程中,优化模型的迭代求解过程需要反复多次调用模拟模型(即对模拟模型进行求解),对于DNAPLs污染含水层多相流数值模拟模型而言,反复多次计算模拟模型会带来庞大的计算负荷,这会严重制约模拟模型和优化模型在DNAPLs污染含水层修复工程实际应用中的可行性。因此,建立合理有效的替代模型成为解决问题的可行途径。替代模型在功能上逼近模拟模型,在计算上则易于解算,大幅度地减少计算负荷。然而,替代模型的研究尚处于尝试探索阶段,其精度的好坏取决于采样方法和替代模型种类的研究选定。因此,本文针对表面活性剂强化的DNAPLs污染含水层修复问题,分别以假想DNAPLs污染含水层和实际污染场地DNAPLs污染含水层为例,开展了多相流模拟模型的替代模型理论和方法的创新性研究。首先在建立多相流数学模拟模型的基础上,分别运用蒙特卡罗采样方法和拉丁对偶变数复合采样方法在多相流模拟模型可控输入变量的可行域内采样,然后运转多相流模拟模型产生输入-输出样品数据集。对比分析了两种方法采样结果的采样效率和覆盖程度,同时为模拟模型的替代模型的建立准备数据样本。然后,基于由两种采样方法和模拟模型获得的输入-输出样品数据集,分别运用双响应面方法和径向基函数人工神经网络方法建立了多相流模拟模型的替代模型。最后,任意选取了一组新的抽注水方案,分别代入到多相流模拟模型和运用不同途径建立的替代模型中求解,并对计算结果进行了对比分析,从中遴选和总结出了合适的建立多相流模拟模型的替代模型的理论和方法。本文的研究取得的主要结论如下:①对于同一种建模方法(双响应面方法或径向基函数人工神经网络方法),基于拉丁对偶变数复合采样建立的替代模型对模拟模型的逼近程度明显高于基于蒙特卡罗采样建立的替代模型对模拟模型的逼近程度。这是由于蒙特卡罗采样法是利用随机数从概率分布中采样的随机采样方法,它得出的样品完全随机出现,常常产生数据点偏聚的问题,抽取出的样品对总体覆盖程度不高;而拉丁对偶变数复合采样属于分层采样,它在保证采样效率的同时,得出的样品更加精确地反映了输入概率函数中值的分布,使样品空间的覆盖程度得到了保证,抽取的样品具有一定代表性。②对于同一种采样方法(蒙特卡罗方法或拉丁对偶变数复合方法),运用径向基函数人工神经网络法建立的替代模型对模拟模型的逼近程度明显高于运用双响应面方法建立的替代模型对模拟模型的逼近程度。这是由于双响应面方法在使用前,都要事先对某问题的输入-输出函数关系类型有一个判断,然后才能确定建立何种形式的回归方程。但经判断后建立的回归方程作为替代模型,其对模拟模型的逼近程度仍然有限。而径向基函数神经网络通过不断调整输入样本的聚类中心和隐含层到输出层之间的权值,使网络的实际输出逐渐向希望输出逼近,最终使其有识别输入模式特征的能力。并且径向基函数人工神经网络收敛速度快,能够找到全局极小。③通过综合对比分析,假想例子与实际例子的计算结果和结论得到了相互印证。运用两种采样方法结合两种建模方法所建立的替代模型在功能上都能逼近模拟模型,均具备了与模拟模型相近的输入输出关系。但是它们对原模拟模型的逼近程度仍有差别,按对模拟模型的逼近程度从低到高排序是:基于蒙特卡罗采样的双响应面模型、基于拉丁对偶变数复合采样的双响应面模型、基于蒙特卡罗采样的径向基函数人工神经网络模型、基于拉丁对偶变数复合采样的径向基函数人工神经网络模型。因此,针对DNAPLs污染含水层修复问题,运用拉丁对偶变数复合采样法结合径向基函数人工神经网络方法建立的多相流模拟模型的替代模型,是更为合理有效的替代模型。更多还原

【Abstract】 Because of the petrol spilling, illegal disposal and contretemps, crude oil and various petroleum products enter into the environment, and cause environmental pollution. Petroleum products are mixtures mainly consist of alkanes, cycloalkanes and aromatic, have carcinogenic, teratogenic and mutagenic potential threats, and are toxic pollutants that are detrimental to human and the environment.The solubility of petroleum contamination is very small with water. The petroleum contamination usually exists in form of non-aqueous phase liquids (NAPLs). DNAPLs have high density, low water solubility and high interfacial tension properties. The remediation of DNAPLs is more difficult than LNAPLs, commonly used out– processing technology is difficult to control it effectively, and the cost of remediation of DNAPLs is very expensive. The cost of a single contaminated site often requires hundreds of millions of dollars. Surfactants flushing technology appearing in recent years is also called Surfactant Enhanced Aquifer Remediation (SEAR), which improves the out– processing technology. Surfactants have solubilization and mobilization for hydrophobic organic pollutants and can improve the solubility and migration of DNAPLs in water. So they allow for more freedom phase DNAPLs into the water and substantially increase the effectiveness of out– processing technology to repair DNAPLs.At present, Surfactant Enhanced Aquifer Remediation (SEAR) is still in development stage and factors effect restoration and repair costs, such as selected positions of pumping and injection wells and the concentration of surfactant are very complex.Therefore, process optimization design of aquifer remediation of contaminated site based on the field investigation, through the rational use of simulation model and optimization model is exigency and has important theoretical and practical significance, which can improve efficiency and reduce the cost of remediation.The application of simulation-optimization approaches for designing the optimal groundwater remediation systems is given more widespread attention. In the process of the use of simulation models and optimization models, the solution procedure of optimization model need repeated call for simulation model what can bring Huge computational burden for multiphase flow numerical simulation model of DNAPLs contaminated aquifer when simulation model calculation. This would seriously restrict the feasibility of the remediation application of simulation model and optimization model in DNAPLs contaminated aquifer. Therefore, the establishment of surrogate model which is reasonable and effective, so that its function can approximate numerical simulation model, and avoid repeated calls for simulation, and to shorten the computing time. It is a feasible way to solve the problem.However, the study of surrogate model is still in the exploratory and attempt stage, its accuracy is good or bad depending on the sampling method and the type of surrogate model.In this study, aiming at the problem of surfactant-enhanced DNAPLs contaminated aquifer remediation, taking the imaginary and real DNAPLs contaminated aquifers as the research object, it made a study of the theory and method of the surrogate model of Multiphase flow simulation model.First, a multiphase flow numerical simulation model of surfactant- enhanced DNAPLs contaminated aquifer was first building as the base. It was used to simulate the migrate law of water, surfactant and DNAPLs. Then study on using Monte Carlo sampling method and Latin antithetic variable composite sampling method for collecting input-output sample data of multiphase flow simulation model. And compared the results of efficiency and sampling coverage of two sampling methods. According to the input - output sample data sets got by two sampling methods, then building surrogate models of multiphase flow simulation model ---dual response surface model and radial basis function artificial neural network model. At last, choosing a new program to testing the level that surrogate model approximate the numerical simulation, and summed up the proper selection of the surrogate model of multiphase flow simulation model.Main conclusions obtained from the paper are as follows:①For the same modeling approach(Dual response surface model and Radial basis function artificial neural network model), the approximation to the simulation model based on Latin antithetic variable composite sampling method higher than Monte Carlo sampling method. This is because the Monte Carlo sampling method is a random sampling method that samples from the probability distribution using random numbers. Its samples appear completely random, which often have the problem of segregation of data points, and the overall coverage of the extracted sample is not high; However, Latin dual variable composite sample is stratified sampling, whose samples reflect the value distribution of the input probability function when ensure the efficiency of sampling. So the coverage of the sample space is guaranteed and good representative samples are taken.②For the same sampling method(Monte Carlo sampling method and Latin antithetic variable composite sampling method), the approximation to the simulation model based on Radial basis function artificial neural network higher than Dual response surface method. This is because before using the dual response surface method, it need to have a judge of the input - output function type of an issue and then to determine what form of regression equation established. However, the regression equation after judged is used as an alternative model, whose approximation of the simulation model is still limited. The radial basis function neural network makes actual output of the network gradually to approach to the desired output by continuously adjust the cluster centers of the input sample and weights between the hidden layers to output layer, and ultimately enable it to identify the characteristics of the input mode. And radial basis function artificial neural network converges fast and can find the global minimum.③Through a comprehensive comparative analysis, the calculation results and conclusions of the imaginary and real examples have been confirmed with each other. The conclusion is that the four surrogate models all can approximate the simulation model, they all possess the similar input and output relationship with simulation model, but the degree of approximation to the simulation model still exist differences. The four surrogate models sorted by the degree of approximation to the simulation model from low to high is, Dual response surface model based on Monte Carlo sampling method, Dual response surface model based on Latin antithetic variable composite sampling method, Radial basis function artificial neural network model based on Monte Carlo sampling method and Radial basis function artificial neural network model based on Latin antithetic variable composite sampling method. Therefore, the final decision of the most suitable surrogate model of remediation of DNAPLs contaminated aquifer is radial basis function artificial neural network model based on Latin antithetic variable composite sampling method.更多还原