Groundwater DNAPL contamination source-sink analysis research system based on the simulation-random statistical method for a practical case

Abstract

Groundwater DNAPL contamination source-sink analysis (GDCSA) has seldom been included in previous research on groundwater DNAPL contamination. In this study, a complete GDCSA research system was first established to include groundwater DNAPL contamination source identification (GDCSI), groundwater DNAPL contamination source-sink response relationship (GDCSRR), and spatial and temporal distribution and quantity (sink) of contamination. The object of most previous GDCSI studies has been a simple hypothetical example in which the contamination source was always generalized as a point. For a complex contamination source that cannot be generalized as a point, identifying its shape depicts the source in actual practical terms and ensures the accuracy of GDCSI. In this case study of a practical dye chemical factory in Hebei Province, China, firstly shape of DNAPL contamination source was identified. Then, GDCSRR was determined based on GDCSI result. Finally, sink of contamination based on GDCSRR was determined. Compared with a single study of GDCSI, this complete GDCSA research system provides a reference for identifying the polluter and also provides a more precise basis for contamination remediation scheme design and contamination risk assessment in practical cases. In this paper, we applied parallel heuristic search iterative process (PHSIP) based on simulation-random statistics method to solve GDCSI. However, the repetitive invocation of numerical simulation model has high computational cost during PHSIP. An effective method is to construct surrogate to emulate the simulation model at low computational cost. However, there is a complex nonlinear mapping relationship between input and output for multiphase flow simulation model with large number of variables and high dimension. The accuracy of a surrogate using shallow learning methods needs to be improved. Therefore, we have introduced deep-belief neural network (DBNN) surrogate to emulate the simulation model. A combination of hypothetical and practical cases was adopted in this study. Test of the proposed approaches for the hypothetical case revealed that the DBNN surrogate approximated the multiphase flow simulation model more closely than shallow learning surrogates, and that the PHSIP obtained accurate identification results. These tested approaches and the GDCSA research system were finally applied to the practical case.