Forecasting groundwater contaminant plume development using statistical and machine learning methods
Date
2022
Authors
McConnnell, Elizabeth, author
Blotevogel, Jens, advisor
Karimi Askarani, Kayvan, committee member
Ham, Jay, committee member
Scalia, Joseph, committee member
Journal Title
Journal ISSN
Volume Title
Abstract
A persistent challenge in predicting the fate and transport of groundwater contaminants is the inherent geologic heterogeneity of the subsurface. Contaminant movement has been primarily modeled by simplifying the geology and accepting assumptions to solve the advection- dispersion-reaction equation. With the large groundwater quality datasets that have been collected for decades at legacy contaminated sites, there is an emerging potential to use data- driven machine learning algorithms to model contaminant plume development and improve site management. However, spatial and temporal data density and quality requirements for accurate plume forecasting have yet to be determined. In this study, extensive historical datasets from groundwater monitoring well samples were initially used with the intent to increase our understanding of complex interrelations between groundwater quality parameters and to build a suitable model for estimating the time to site closure. After correlation analyses applied to the entire datasets did not reveal compelling correlation coefficients, likely due to poor data quality from integrated well samples, the initial task was reversed to determine how many data are needed for accurate groundwater plume forecasting. A reactive transport model for a focus area downgradient of a zero-valent iron permeable reactive barrier was developed to generate a detailed, synthetic carbon tetrachloride concentration dataset that was input to two forecasting models, Prophet and the damped Holt's method. By increasing the temporal sampling schedule from the industry norm of quarterly to monthly, the plume development forecasts improved such that times to site closure were accurately predicted. For wells with declining contaminant concentrations, the damped Holt's method achieved more accurate forecasts than Prophet. However, only Prophet allows for the inclusion of exogenous regressors such as temporal concentration changes in upgradient wells, enabling the predictions of future declining trends in wells with still increasing contaminant concentrations. The value of machine learning models for contaminant fate and transport prediction is increasingly apparent, but changes in groundwater sampling will be required to take full advantage of data-driven contaminant plume forecasting. As the quantity and quality of data collection increases, aided by sensors and automated sampling, these tools will become an integral part of contaminated site management. Spatial high-resolution data, for instance from multi-level samplers, have previously transformed our understanding of contaminant fate and transport in the subsurface, and improved our ability to manage sites. The collection of temporal high-resolution data will similarly revolutionize our ability to forecast contaminant plume behavior.