Identification Of Groundwater DNAPLs Contamination Source Based On Combined Heuristic Search Method

Groundwater pollution is often characterized by concealment and delay in discovery,which leads to the lack of understanding and mastery of the status of pollution sources in underground aquifers,including the location,distribution pattern and release history of pollution sources.This brings great difficulties to the reasonable design of groundwater pollution remediation scheme,accurate identification of pollution responsibility and pollution risk warning.Therefore,the identification of groundwater pollution sources has important practical needs.Groundwater pollution source identification refers to the inversion solution of mathematical simulation model describing groundwater pollution based on actual monitoring data of groundwater pollution(monitoring data of water level and pollutant concentration),as well as auxiliary information such as field investigation and professional experience,so as to identify and determine the status of DNAPLs pollution sources in underground aquifers.Identification of groundwater pollution source belongs to mathematical physics inverse problem or mathematical physics inversion problem in mathematics.Most of the inverse problems have the characteristics of unsuitability and nonlinear,which is also the difficulty of the inverse problem.Random statistical method and optimization method are two main methods of groundwater pollution source identification,and one of them is usually used in previous studies.Each approach has its own advantages and limitations.Among them,the optimization method has a strong fine search ability,but it still depends on the selection of initial value.If the initial value is far from the true value,the convergence speed will be very slow,or it is easy to fall into premature convergence,and it is difficult to obtain the optimal solution.Random statistical method can provide not only point estimation but also interval estimation and probability distribution estimation for variables to be identified.However,in the face of more complex problems,the precision of random X statistical method is relatively weak,and the identification accuracy needs to be improved.In order to learn from each other,this paper explores the combined application of stochastic statistical method and optimization method,and develops a combined heuristic search method to identify groundwater pollution sources by complementing the advantages of the two methods.On the one hand,point estimation and interval estimation provided by random statistical method are used to provide better initial value and initial estimation interval for optimization method.On the other hand,the optimization method(establishing and solving the optimization model of pollution source identification)is used to enhance the fine search ability,obtain the global optimal solution with high precision,and improve the identification accuracy.In this paper,through the cross and fusion application of mathematical physics forward and inversion,deep learning,combined heuristic search method and other theoretical methods,the pending scientific problems in the forefront of pollution source identification of Dense non-aqueous phase fluid(DNAPLs)are studied.In the practical problem of groundwater pollution source identification,the pollution source can often be generalized as a point(point distribution).However,some practical problems are not the same.It is difficult to generalize the distribution of pollution sources into points.If some pollution sources have a long linear distribution,it will be seriously divorced from reality if they are still treated as points.When faced with this situation,it is necessary to identify the distribution patterns of pollution sources in the underground aquifer.In this paper,a dye chemical pollution site is taken as a case study area.Based on the field investigation and data analysis,it is difficult to generalize the distribution pattern of pollution sources into a single point,so the identification of the distribution pattern of pollution sources must be considered.Therefore,in this case,the location,distribution pattern and release history of DNAPLs pollution sources in underground aquifers and parameters of multiphase flow simulation model were jointly identified in this paper.On the basis of field investigation and data analysis,the hydrogeological conditions of the study area are qualitatively analyzed and generalized,and the conceptual model of the study area is established.Based on the conceptual model,a numerical simulation model of multiphase flow polluted by DNAPLs was established.Then,the sensitivity analysis was carried out on the parameters of the multiphase flow simulation model,and the parameters that have a significant impact on the output of the model were selected as the parameters to be identified,and the parameters to be identified together with the status of the pollution source were taken as the variables to be identified.In the search iterative process of solving groundwater pollution source identification problem,it is necessary to repeatedly call the numerical simulation model of solving multiphase flow,which will bring huge calculation load and long calculation time.In this paper,an alternative model for multiphase flow simulation is established to solve this problem.The alternative model can not only greatly reduce the computational load caused by repeated calls to the simulation model,but also keep better computational accuracy.The Latin hypercube sampling method is used to sample the variables to be identified in the feasible region,and the sampling results are input to the multiphase flow simulation model for forward calculation,to obtain the input-output sample data of the multiphase flow simulation model,which can be used as training samples and test samples for the establishment of alternative models.Using gaussian process method,kernel extreme learning machine method,deep confidence neural network method and stack autoencoder neural network method respectively,the gauss process replacement model,kernel extreme learning machine replacement model,deep confidence neural network replacement model and stack autoencoder replacement model of multiphase flow simulation model are established.Four alternative models were trained using training samples.Using test samples,the approximation accuracy of the four alternative models to the multiphase flow simulation model was analyzed,and the advantages and applicability of the stack autoencoder alternative model were compared and analyzed.In this paper,a combined heuristic search method is developed by combining stochastic statistics method and optimization method,which is used in the research of groundwater pollution source inversion identification.XIIFirstly,the heuristic search iterative process is designed and constructed by applying the theory and method of random statistics.A state evaluation function based on Bayesian formula and stack autoencoder alternative model is established.The variable radius free search method is applied to select and determine the trial points of each iteration.For each iteration,the judgment is made according to the calculation results of Tsallis formula based on state evaluation function.If the condition of state transition is met,the state transition is carried out,and the trial point of this iteration is taken as the starting point of the next iteration.If not,the variable radius free search method should be used to re-select the trial point of this iteration.When the convergence condition is reached,the point estimation,interval estimation and probability distribution estimation of the variable to be identified are obtained.The point estimation and interval estimation provided by random statistical method are used to provide better initial value and initial interval for optimization method.According to the interval estimation and probability distribution estimation provided by the random statistical method,the training samples and test samples of the stack autoencoder alternative model were extracted again,and the stack autoencoder alternative model with better accuracy was reconstructed,which was embedded into the optimization model as equality constraints.Then,the optimization model of groundwater DNAPLs pollution source identification is established.In order to improve the solution accuracy of the optimization model,the variable probability migration strategy and the information sharing mechanism in the particle swarm optimization algorithm were introduced to improve the colony foraging optimization algorithm,and the variable probability hybrid particle swarm optimization algorithm was constructed.Finally,the variable probability mixed particle swarm foraging optimization algorithm was used to solve the optimization model,and the status of groundwater DNAPLs pollution source and the identification results of the multi-phase flow simulation model parameters were obtained.Through the above research,the following main conclusions are drawn:(1)In this paper,the status of DNAPLs pollution sources in an underground aquifer in a case study area(including the location,distribution form and release history of pollution sources)and the parameters in the simulation model are jointly identified.For the complex pollution sources that cannot be generalized as points,this paper carried out the identification of the distribution patterns of pollution sources,which can depict the real distribution patterns of actual pollution sources and make the identification results of the distribution patterns of pollution sources more practical.(2)In this paper,gaussian process method,kernel extreme learning machine method,deep confidence neural network method and stack autoencoder neural network method are respectively used to establish the alternative model of multiphase flow simulation model.Through the analysis and verification of different modeling methods for the alternative models,the results show that the approximation accuracy of the stack autoencoder alternative model based on deep learning and the deep confidence neural network alternative model is significantly higher than the Gaussian process alternative model based on shallow learning and the kernel extreme learning machine alternative model.Among them,the stack autoencoder substitute model has the highest approximation accuracy to the multiphase flow simulation model,and can better fit the complex nonlinear mapping relationship between the inputs and outputs of the multiphase flow simulation model.(3)In this paper,based on the combined use of random statistical method and optimization method,the combination of the two methods is applied,a combination of heuristic search method is developed,which can make full use of the random statistic method provides the point estimation and interval estimation,in order to optimize the method provides a better initial value and the estimated interval,and to use optimization methods to enhance the capacity of the fine search,to improve search effect and improve the recognition accuracy.(4)In the random statistical method,this study adopts two improved techniques to improve the search effect.First,the variable radius free search method is applied to select and determine the trial point of each search iteration,and the ergodicity and efficiency of search are taken into account by constantly adjusting the size of search radius.The second is to construct Tsallis formula based on state evaluation function,and take its calculation results as the discrimination basis of state transition.By adjusting the control parameters of Tsallis formula,the search process can find a balance between ergodicity and efficiency.(5)In order to improve the precision of the optimization model,this study explore the introduction of variable probability migration strategy and information sharing mechanism of particle swarm optimization algorithm,the bacteria foraging optimization algorithm is improved,and build a variable probability-particle swarm algorithm-bacteria foraging optimization algorithm,can avoid falling into local minima,and can search to the global optimal solution quickly,improve the accuracy of solving the optimization model.To sum up,this paper carried out innovative research on groundwater DNAPLs pollution source identification,which enriched the theoretical basis of groundwater pollution source identification and provided technical support for the practical application of groundwater pollution remediation projects.