Forecasting Groundwater Contaminant Plume Development Using Statistical and Machine Learning Methods

Abstract

Forecasting groundwater contaminant plume development is critical for determining risks to downgradient receptors and predicting the time to site closure. However, accurate forecasts are a formidable challenge due to the complexities of a heterogeneous subsurface. While historically groundwater well data in combination with numerical flow models have been used for this task, the advent of machine learning offers new data-driven opportunities for improving contaminant fate and transport predictions. In this study, we interrogate the viability of two forecasting models—Prophet and damped Holt's exponential smoothing model—for predicting groundwater contaminant plume development. The impacts of spatial and temporal data density on the accuracy of the forecasts are evaluated. For wells with declining contaminant concentrations, the damped Holt's method achieves more accurate forecasts. However, only Prophet allows for the inclusion of exogenous regressors, enabling predictions of future declining trends in wells with still increasing contaminant concentrations. Application of these models does not only require robust training data, but also an understanding of model biases. Overall, powerful data-driven models are already available for contaminant plume prediction, but groundwater sampling approaches will have to improve, for instance, through the collection of real-time spatial and temporal high-resolution data, to take full advantage of their capabilities.