Applicability of Advanced Forecasting in Subsurface Remediation

Using forecasting models to help manage contaminated sites has large potential, but also faces many obstacles. Exploiting historical data from field and laboratory analysis of groundwater monitoring well samples, we attempt to increase our understanding of complex interrelations and to build a suitable model for estimating the time to site closure. Preparing a model to predict subsurface contaminant concentrations or rates requires thoughtful investigation of the controlling factors in a dataset. This includes handling non-detect values, joining and reshaping data, and exploratory analysis. Our goal is to investigate large legacy site datasets to determine how accurately we can apply forecasting models to historical data and improve site management. Preliminary investigations of historical groundwater datasets included correlation analyses to determine the factors that have the greatest prediction power. After basic correlation analysis applied to datasets for entire sites did not reveal compelling evidence, we chose to focus on smaller remediation areas as well as using alternative methods for exploratory analysis. Finding these patterns will allow us to choose which parameters are most relevant to include as predictors, and which can potentially be eliminated from long-term groundwater monitoring programs. Additionally, we seek to understand the data needs for a functional machine learning model. If the models we want to use are not sufficient to meet our goals using existing data, what is the density of data that would be required to make them useful? We are working to use a numerical groundwater flow and transport model to create a synthetic dataset, which we can manipulate to determine the amount of data required to make stable predictions (Figure 1). This information can be applied to sites to establish sampling frequency and sensor deployment needs. Spatial high-resolution data have previously transformed our understanding of contaminant fate and transport in the subsurface, and improved our ability to manage sites. The collection of temporal and spatial high-resolution data will similarly revolutionize our ability to forecast contaminant concentrations and the time to closure. With the rapid emergence of sensor-based collection of better and cheaper data, this revolution is imminent.