Spatial-temporal analysis of groundwater well features from neural network prediction of hexavalent chromium concentration

Abstract

Groundwater monitoring is a crucial part of groundwater remediation that produces data from various strategically placed wells to maintain a water quality standard. Using the United States Department of Energy’s Hanford 100-HRD area well data, recurrent neural networks are trained in the form of one-dimensional Convolutional Neural Networks (CNNs), Long Short Term Memory (LSTM) networks, and Dual-stage Attention-based LSTM (DA-LSTM) networks to reduce monitoring costs and increase data sampling responsiveness that is subject to laboratory analysis delays, with the best network being DA-LSTM achieving an R2 score of 0.82. There is also the provision of novel well relation analysis through two method of explanation coming from SHapley Additive exPlanations (SHAP) and attention mechanisms within DA-LSTM. Extracting feature well importance from these two methods and using target wells as vehicles for pipeline demonstration, it is found that wells 199-D5-160, 199-D4-68, 199-D2-11, 199-D8-94, and 199-D2-10 have had the most year-long impact on Hexavalent Chromium (CrVI) concentration at the target during the testing year, while wells 199-D8-94, 199-D4-95, 199-D8-53, 199-D4-85, and 199-D4-99 are found to be strong candidates affecting target CrVI instantaneous change during a particular timeframe.