Performance of Regression and Artificial Neural Network Models, Informed with an In Situ Sensor Network, in Forecasting Groundwater Abstraction in the Central Valley, California

Prolonged drought conditions in the western United States have spurred the passage of water-related legislation, including laws to protect groundwater aquifers by reducing or capping groundwater abstraction. In California, the Sustainable Groundwater Management Act was passed in 2014 with the goal of preserving groundwater resources in the Central Valley Aquifer. While there are a plethora of groundwater level and hydrologic data that can be used for regional groundwater planning, water managers and stakeholders often do not have data on the timing, location, and quantity of groundwater abstraction. In this analysis, we use an in situ groundwater abstraction data set, derived from electrical sensors, coupled with hydrologic and climatic data to create a groundwater abstraction forecast model. We compare the performance of the four model algorithms to determine a suitable model for this application. Artificial neural network models outperform multiple-linear regression and multiple-logistic regression models, with the best-performing model generating an R-2 of similar to 0.90. An analysis of the spatial and temporal variability in model error shows that the model performs best on a monthly time step, forecasts are reliable within a two-month lead time, and regional heterogeneity may be a source of error within the model. These results lay the groundwork for a groundwater abstraction forecast tool that may aid water managers in groundwater policy compliance.