Study On Source Identification And Source-sink Relationship Of LNAPLs Contamination In Groundwater

In recent years,oil and its derivatives often leak into the underground aquifer and pollute the groundwater due to sudden accidents or improper disposal.The solubility of petroleum organic pollutants in water is generally small,and they usually exist in the underground aquifer in the form of non-aqueous phase liquids（NAPLs）.According to the different density,NAPLs can be divided into two types:one is dense non-aqueous phase liquids（DNAPLs）with density greater than water,and the other is light non-aqueous phase liquids（LNAPLs）with density less than water.NAPLs are often highly toxic.If they are not repaired and treated in time,they will cause serious pollution to groundwater quality.However,the premise of effective groundwater NAPLs pollution remediation lies in the accurate determination of pollution source information（number,location and release history of pollution sources）.If it is treated blindly without understanding and mastering the pollution source information,it is often difficult to achieve satisfactory remediation results.Therefore,before pollution remediation,it is particularly important to carry out groundwater pollution source identification research,and then understand and master the relevant information of pollution sources.Groundwater pollution source identification refers to the inversion solution of the simulation model describing the migration law of pollutants in groundwater through the auxiliary information obtained through field investigation and data collection,combined with the limited field monitoring data（including water level and water quality monitoring data）,to deduce the information of pollution sources,including the number,location and release history of pollution sources（i.e.the initial release time of pollutants,the termination release time of pollutants and the change of release intensity with time）.Because the research of groundwater pollution source identification is of great significance,it has attracted extensive attention in recent years.However,at present,the research on the source identification of groundwater LNAPLs pollution is still scarce,and there is an urgent need to carry out relevant research work.In addition,there is a strong causal relationship between pollution sources（the source of pollutants）and pollution sinks（the destination of pollutants）.Therefore,it is very necessary to consider the two at the same time and make an in-depth study on the causal relationship between them.However,at present,the research on the source-sink relationship of groundwater pollution is very scarce,and the research on the source-sink relationship of groundwater LNAPLs pollution is rarely reported.Therefore,there is an urgent need to focus on the unsolved problems in the research frontier of the source-sink relationship of groundwater LNAPLs pollution.In view of the above problems,this study comprehensively applies many theories and methods,such as multiphase flow numerical simulation model,biodegradation numerical simulation model,deep residual network method based on Resnet-18structure,variable density grid search strategy,Adaptive cyclic improved iterative process,Ensemble smoother-optimal differential evolution Markov chain(ES-DEMC_（B）)algorithm,Monte Carlo method and so on,Firstly,the identification of groundwater LNAPLs pollution source is studied.On the basis of identifying and determining the relevant information of pollution sources,the relationship between groundwater LNAPLs pollution source and sink is studied.On the basis of field investigation and data analysis,the geological and hydrogeological conditions of the polluted site are generalized,and the conceptual model of the polluted site is established.According to the data and professional experience,the parameters and source information of groundwater LNAPLs pollution multiphase flow and biodegradation numerical simulation models were given the initial estimation and the initial estimation of the value range,and the multiphase flow and biodegradation numerical simulation models that could depict the LNAPLs pollutant migration rule and biodegradation process was established.Through the sensitivity analysis of the parameters in the simulation model,the more sensitive simulation model parameters are selected,which together with the pollution source information are used as the unknown variables in the following study of groundwater pollution source identification.Because iterative operation is often involved in the research of groundwater LNAPLs pollution source identification,the simulation model needs to be called repeatedly,which will bring huge calculation load and lengthy calculation time.This study solves this problem by establishing a surrogate model of multiphase flow and biodegradation numerical simulation models.The surrogate model can not only greatly reduce the calculation load and calculation time,but also maintain good approximation accuracy.Kriging method,support vector regression method,deep belief network method and deep residual network method based on Resnet-18 structure（Resnet18-DRN）are used to establish surrogate models for multiphase flow and biodegradation numerical simulation models.Using training samples,the above four surrogate models are trained.Using the test samples,the approximation accuracy of the four surrogate models to the simulation model is analyzed,and the advantages and applicability of the deep residual network surrogate model based on Resnet-18 structure are compared and analyzed.In addition,to further improve the approximation accuracy of the Resnet18-DRN surrogate model to the simulation model,the variable density grid search strategy is applied to optimize the hiperparameters in the deep residual network（that is,the parameters that need to be set before establishing the surrogate model,rather than those obtained through training）.In addition,the quality of training samples is an important factor to determine the accuracy of surrogate models.In the past,the training samples of the surrogate model are often generated at one time,which makes the quality of the training samples difficult to be continuously improved with the iteration.Therefore,it is difficult to establish a high-precision surrogate model,which will affect the solution accuracy of groundwater LNAPLs pollution source identification problem.Therefore,to improve the quality of training samples and improve the solution accuracy of groundwater LNAPLs pollution source identification problem,this study constructs an adaptive cyclic improved iterative process based on deep learning.Firstly,according to the input-output samples（training samples）of multiphase flow and biodegradation numerical simulation model,the surrogate model of multiphase flow and biodegradation numerical simulation model is established by using Resnet18-DRN.Based on the established surrogate model,the ensemble smoother algorithm is applied to obtain the identification results of the variables to be solved.Then,try to use the identification results of the variables to be solved to spit out the old and absorb the new for the previous training samples,absorb the high-quality new training sample,and abandon the poor old training sample.According to the improved and updated training samples,a new surrogate model with better quality is established.Based on the new surrogate model,the ensemble smoother algorithm is applied to obtain new and more accurate identification results of the variables to be solved.The above process is repeated until the convergence conditions are met,and the iterative process is terminated.Through the above adaptive cyclic improved iterative process based on deep learning,the training samples,surrogate models and identification results of variables to be solved can be improved and updated together with the iterative process,to improve the solution effect of groundwater LNAPLs pollution source identification problem.There are many methods to solve the problem of groundwater pollution source identification.In the previous research,a single method is often used for source identification.However,each method has its own advantages and disadvantages.Among them,Kalman filter method has small amount of calculation and fast running speed,and can quickly obtain the identification results.However,when the nonlinear degree of the problem is strong,its optimization ability is relatively weak,and the accuracy of the identification results is not high.The optimization ability of stochastic statistical method is less affected by the nonlinearity of the problem and has strong optimization ability.However,if the initial value is far from the true value,the convergence speed will be very slow and the amount of calculation will be large.Therefore,to learn from each other,this study combines the ensemble smoother algorithm（belonging to Kalman filter method）and the optimal differential evolution Markov chain algorithm（belonging to stochastic statistical method）to construct the ES-DEMC_（B）to solve the problem of groundwater LNAPLs pollution source identification.Firstly,the identification results of the ensemble smoother algorithm are used to provide better initial values for the optimal differential evolution Markov chain algorithm,which can speed up the convergence speed,avoid local extreme points and quickly find the global optimum.Next,the optimal differential evolution Markov chain algorithm is used to identify the pollution source of groundwater LNAPLs,and the final identification results of each variable to be solved（a posteriori probability distribution and point estimations）are obtained.The ES-DEMC_（B）algorithm is applied to a hypothetical example and a real example respectively to analyze the effectiveness of the algorithm.Through the above research on source identification of groundwater LNAPLs pollution,the identification results of pollution source information（horizontal coordinates and longitudinal coordinates of pollution source location,initial release time of pollutants,termination release time of pollutants and release intensity of pollution sources）and simulation model parameters are obtained.Then,the pollution source sink relationship of groundwater LNAPLs is analyzed through deterministic simulation and stochastic simulation.Firstly,carry out deterministic simulation,substitute the above identified pollution source information and model parameter results into the multiphase flow and biodegradation numerical simulation models for solution calculation,and quantitatively analyze the destination of pollutants after entering the underground aquifer,that is,calculate the distribution of pollutants at each destination（i.e.the residual amount of pollutants,biodegradation amount of pollutants and boundary outflow amount of pollutants in the simulation area in the whole simulation period）.Then,taking the pollution source information as random variables and the model parameters as known constants,Monte Carlo stochastic simulation is carried out.According to the posterior probability distribution of the obtained pollution source information,multiple groups of random samples are generated.The random samples are substituted into the multiphase flow and biodegradation numerical simulation models for solution and calculation,and the distribution of each destination of the corresponding pollutants is obtained.Through statistical analysis of multiple groups of results,the probability distribution of each disposal and allocation of pollutants is obtained,and the uncertain impact of the random change of pollution source information on the distribution of pollutants at each destination is analyzed.Through the above research,this paper draws the following conclusions:（1）In this study,Kriging method,support vector regression method,deep belief network method and deep residual network method based on resnet-18 structure are used to establish surroagte models for multiphase flow and biodegradation numerical simulation models,and the accuracy of the four surrogate models is compared and analyzed,The results show that the accuracy of the two deep learning surrogate models（deep belief network surrogate model and deep residual network surrogate model based on Resnet-18 structure）is significantly better than that of the two shallow learning surrogate models（Kriging surrogate model and support vector regression surrogate model）,and among the two deep learning surrogate models,the accuracy of the deep residual network substitution model based on Resnet-18structure is higher.In addition,by applying the variable density grid search strategy to optimize the hyperparameters in the deep residual network surrogate model based on Resnet-18 structure,compared with the conventional grid search strategy,it can further improve the approximation accuracy of the deep residual network surrogate model based on Resnet-18 structure to the simulation model.（2）In the research of groundwater LNAPLs pollution traceability identification,the research and construction of adaptive cyclic improved iterative process based on deep learning can significantly improve the accuracy of groundwater LNAPLs pollution source identification.In the study of the hypothetical example,adaptive cyclic improved iterative process is carried out for ten rounds,and the average relative error is reduced from the initial 13.11%to 5.54%,and the maximum relative error is reduced from 41.37%to 14.36%.In the study of the real example,adaptive cyclic improved iterative process is carried out for 12 rounds,and the identification results are evaluated by calculating the proximity between the output value and the field monitoring value.The results show that the certainty coefficient R² between the two is increased from 0.6722 to 0.8311.Therefore,in both hypothetical and practical examples,by using adaptive cyclic improved iterative process based on deep learning t,the accuracy of groundwater LNAPLs pollution traceability identification results has been significantly improved.（3）In this study,the ES-DEMC_（B）is constructed by combining the ensemble smoother algorithm and the optimal differential evolution Markov chain algorithm,which can make full use of the advantages of the fast running speed of the ensemble smoother algorithm,provide a better initial value for the optimal differential evolution Markov chain algorithm,and make it converge with fewer iterations,It can also give play to the strong optimization ability of the optimal differential evolution Markov chain algorithm and improve the accuracy of groundwater LNAPLs pollution source identification results.（4）Based on the numerical simulation model of multiphase flow and biodegradation,this study analyzes the pollution source-sink relationship of groundwater LNAPLs through deterministic simulation and stochastic simulation.It can be concluded that most of the pollutants entering the groundwater aquifer are still left in the aquifer of the simulation area,some are biodegradable,and a small part flows out from the boundary of the simulation area.When considering the random change of pollution source information,the residue of pollutants in the simulation area obeys the normal distribution,and the biodegradation and boundary outflow of pollutants in the simulation area do not obey the normal distribution.In addition,the correlation between the information of pollution sources and the assigned number of pollutants can also be obtained.