Predicting Daily Groundwater Levels with Deep Learning Models

Physics based groundwater simulation models typically require high compute times, input parameter calibration and sensitivity analysis before being applicable for groundwater management in a particular watershed. In order to make informed management decisions, simulations must be run for different future climate and human interference scenarios to assess the benefits of water use and the associated risk of groundwater shortages. The computational cost of the simulations impedes their use for making timely groundwater decisions.

Deep learning (DL) models provide a computationally efficient alternative to mitigate the computational costs from simulations. These models are able to capture the spatial and temporal interactions in a watershed. We train our DL models on daily historical observations of temperature, precipitation, and groundwater levels. The DL models are computationally lightweight and can be optimized and trained in a few minutes on a laptop, as opposed to simulation models that may require many hours of compute time on a high performance computer. The trained DL models allow us to make daily groundwater level predictions within seconds. We show the performance of two DL models, namely Long Short Term Memory Recurrent Neural Networks and Convolutional Neural Networks for predicting the daily groundwater levels in Butte County, CA, and Rifle, CO. The groundwater level predictions are informed by future predictions of temperature and rainfall that we obtain from future climate simulations. We find that our optimized DL models accurately predict the future groundwater levels, which will enable groundwater sustainability agencies to make informed and timely decisions for groundwater management.