Hossein Azhdari Bajestani; Saeed Alimohammadi
Abstract
Introduction: Dense Nonaqueous Phase Liquids (DNAPL) are the most common types of groundwater pollution. Surfectant Enhanced Aquifer Remediation (SEAR) is one of the most common methods of DNAPL-contaminated aquifer remediation. Due to the high cost of the chemicals used in this method (surfectants or ...
Read More
Introduction: Dense Nonaqueous Phase Liquids (DNAPL) are the most common types of groundwater pollution. Surfectant Enhanced Aquifer Remediation (SEAR) is one of the most common methods of DNAPL-contaminated aquifer remediation. Due to the high cost of the chemicals used in this method (surfectants or cosolvents), it is necessary to choose the appropriate wells pattern, and the optimal pumping rates. UTCHEM simulation software has the ability to model the fate and transport of DNAPL and the application of the SEAR method. The main problem with this software is the long time required to run multiple senarioes when using optimization algorithms. The purpose of this study is to use two machine learning methods (Artificial Neural Network and K nearest neighbor) as sorrugate simulation model and imbedding the best one into the LINGO software to optimize the SEAR method. Material and methods: in the implementation of The SEAR method, the quantitative and qualitative aquifer data are required to model how to diffuse, transmit and delete DNAPLs in UTCHEM software. For this purpose information from the Camp Lejeune site in North Carolina, USA were used. In this study, by examining a variety of alternative models based on machine learning methods and implementing 250 different scenarios in UTCHEM software, two models, Artificial Neural Network (ANN) method, and k-nearest neighbors (KNN) were used to simulate the SEAR method and developing alternative model. In order to validate the two alternative models, 50 new scenarios were implemented in UTCHEM software and their percentage of removal was obtained. Also, using two alternative models, the percentage of removal of 50 scenarios were determined. in order to evaluate the performance of alternative models , the root mean square error (RMSE) was used and was compared with the results of other researchs. Finally an alternative model with more accuracy was used in LINGO software to optimize the Surfactant Enhanced Aquifer Remediation method (SEAR).Results and discussion: RMSE values in the results obtained from alternative models ANN and KNN in the validation stage were 0.67 and 1.66 respectively, which indicates the high accuracy of both alternative models, especially ANN. The average run time of each UTCHEM software in this study was 45 minutes, while in the alternative model it was reduced to a few seconds; LINGO software also examined about 21,500 different scenarios in 30 minutes to determine the optimal scenario, while the time required for this task is more than 16,000 hours if the alternative model is not used. Based on the position and discharge of active wells in the optimum scenario, it was found that firstly the existing wells upstream and downstream of the Pollution have the most impact on the remediation and secondly, the time factor is more effective than the wells pumping discharge in the remediation. The optimized scenario obtained in this study remediates the DNAPL-contaminated area by up to 95% at a lower cost than the costs reported in the Camp Lejeune project over a period of 30 days.Conclusion: Based on the results obtained in this study, it was found that the use of machine learning algorithms such as ANN and KNN, along with LINGO optimization software, which is one of the most powerful software for solving linear and nonlinear optimization problems, in addition to having the right accuracy, significantly reduces the time required to find the optimal scenario.
