Optimization Design And Uncertainty Analysis Of DNAPLs-contaminated Groundwater Remediation

Petroleum leakage during exploitation,storage and transportation can result in groundwater pollution,endangering ecological environment and the safety of drinking water.Petroleum pollutants,after flowing into the groundwater,exist in the form of non-aqueous phase liquids(NAPLs).NAPLs,according to the difference in density,can be divided into two categories:light non-aqueous phase liquid(LNAPLs),with the density smaller than water,and dense non-aqueous phase liquid(DNAPLs),with the density greater than water.DNAPLs-contaminated sites are more difficult to remediate than LNAPLs,because DNAPLs exhibit properties such as high density,low solubility,and high interfacial tension.The conventional pump-and-treat technique is relatively ineffective in removing DNAPLs from the subsurface.Surfactant enhanced aquifer remediation(SEAR)is an improvement on the pump-and-treatment technique,with the addition of surfactants to effectively raise the level of DNAPLs solubility and mobility in groundwater,and to elevate the remediation efficiency considerably of DNAPLs-contaminated groundwater by the pump-and-treatment technique.However,use of SEAR is very expensive.Several factors such as the locations of injection and extraction wells,the rates of injection and extraction wells,and surfactant concentration affect the remediation cost and efficiency significantly.In order to elevate the restoration efficacy and save cost,it is necessary to construct and solve a simulation model and optimization model for analysis and optimization of multiple remediation strategies.Besides,it is also necessary to build a surrogate model to replace the simulation model so as to reduce the enormous computational burden and time resulting from repeated and multiple computation of the simulation model in course of optimization solution.The surrogate model is an approximation of the simulation model,and it can be used to markedly reduce computational burden and time,and a satisfactory level of accuracy for computation is maintained.However,some degree of uncertainty exists for both the simulation model and the surrogate model;if it is overlooked,it will prove difficult to conduct a reliability analysis of the best remediation strategy obtained.For the remediation of groundwater contaminated by nitrobenzene,a type of DNAPLs,at a chemical plant,this paper has analyzed and optimized remediation strategies and conducted a reliability analysis of the optimal remediation strategy,by the joint use of multiphase flow simulation model,adaptive updating sampling,surrogate modeling,mixed integer nonlinear programming optimization model and uncertainty analysis.First,on the basis of building a multiphase flow simulation model,a choice of Latin hypercube sampling method was made as an initial sampling method for adaptive updating sampling;sampling was completed within the range of controllable input variables in the simulation model,and output data sets were obtained by operating the multiphase flow simulation model.Later,a surrogate model of the multiphase flow simulation model was constructed by kernel extreme learning machine method,Kriging method and support vector regression method.On that basis,the construction of an ensemble surrogate model made up of multiple individual surrogate models was completed;through analyzing and comparing the closeness of approximation of different surrogate models to the simulation model,the surrogate model with the greatest accuracy was singled out.Then,0-1 mixed integer nonlinear programming model was constructed,with the locations of pumping and injection wells(0-1 integer variables),the amount of pumped and injected water and remediation time(continuous variables)as decision variables,and with the surrogate model and other constraints as constraint conditions,and the model was solved by means of genetic algorithm.Starting from the second round of the cycle,insertion of the current optimum solution into the original training sample was completed,together with elimination of the sample with the greatest difference from the current optimum solution;reconstruction of a surrogate model and optimization model was completed by using the updated sample,and a new optimum solution was obtained by solving the model.Iteration was repeated until satisfaction of convergence conditions.The final round of the solution worked out was the deterministic optimized remediation strategy.Then,on the basis of treatment of the correlation between simulation model parameters by Copula function,an uncertainty analysis of the simulation model was conducted by using Monte Carlo simulation.Then,calculation and counting of the residual between the simulation model and the surrogate model were completed,with an uncertainty analysis of the surrogate model.Finally,stochastic nonlinear programming optimization model was built by considering the uncertainty of both the simulation model and the surrogate model,and the model was worked out by deterministic constraint equivalent method and genetic algorithm to obtain the optimal remediation strategies at different confidence levels.It has enriched and expanded relevant theories and technological implications of optimization process of remediating DNAPLs-contaminated aquifer.Through the above research,the following conclusions are drawn:(1)The design and application of adaptive updating sampling have repeatedly updated and ameliorated training samples,the surrogate model and optimization model in the process of feedback correction.In the process of adaptive feedback correction,the total number of training samples is unchanged,without any additional computational burden.(2)In comparison with Kriging surrogate model and support vector regression surrogate model,kernel extreme learning machine surrogate model is the closest approximation to the simulation model.The ensemble surrogate model of Kriging-Kernel extreme learning machine is better than any single surrogate model in closeness,which further improves the approximation of the surrogate model to the simulation model.It indicates that the ensemble surrogate model mixes all the strengths of single substitute models so as to be a more effective method of surrogate modeling.Therefore,the ensemble surrogate model of Kriging-Kernel extreme learning machine is embedded into the optimization model.(3)The Copula function is effective in dealing with the correlation between simulation model parameters,which makes the uncertainty analysis of the simulation model more in line with objective reality.(4)Building and solving 0-1 mixed integer nonlinear programming optimization model has effectively solved the issue of concurrent optimization of the location of pumping-injection wells(0-1 integer variables),the amount of pumped and injected water(continuous variables)and remediation time(continuous variables),(5)Stochastic nonlinear programming optimization model is built by considering at the same time the uncertainty of both the simulation model and the surrogate model.The optimum remediation strategies are obtained at different confidence levels by working out and optimizing the model.With the conditions for remediation objective(removal rate)constant,the higher the confidence level of remediation strategy,the more expensive the remediation.Therefore,decision makers can weigh remediation costs and remediation strategies at different confidence levels and make a choice in the light of the actualities.