Evaluation of Input Variables for Neural Network Models used in Groundwater Level Forecasting for Sunflower County, Mississippi

Author(s): Guzman, S.; Paz, J.; Tagert, M.

Declining water levels in the Mississippi River Alluvial Aquifer (MRVA) are due to the expansion of irrigated acreage and increasing water demand in the Mississippi Delta region, causing the need to develop forecasting tools and improve conservation measures. One of the tools explored in recent investigations is the Artificial Neural Network (ANN) that has grown in popularity in terms of its application in modeling and forecasting nonlinear hydrologic time series such as groundwater levels. For instance, a previous study demonstrated that an ANN with 2 hidden layers, 100 time delays and Bayesian Regularization training algorithm had the best model architecture that provided predictions of daily groundwater levels up to three months ahead. The effectiveness of ANN in forecasting daily groundwater levels depends on different input datasets as well as on the network learning capacity. An important step in the ANN development process is the evaluation of significant input variables, given that not all of them are powerful predictors of the model output. In this study, the performance of an ANN trained with a Bayesian Regularization algorithm and different input variable combinations was evaluated to determine the optimal model that can simulate groundwater trends up to three months in a USGS monitoring well located in Sunflower County, Mississippi. Nine years of daily groundwater level measurements were collected and partitioned into training and validation data sets. At the same time, input time series such as daily evapotranspiration rates, calculated by the Priestly-Taylor method, and daily precipitation were also partitioned into training and validation sets. The evaluation of the model performance under different input variables was based on the Mean Square Error (MSE) and correlation statistic estimations. The use of ANN with significant input variables provides useful information for the management of water withdrawals either per well or on a regional level in order to implement different conservation practices.

Go back


Past Conference Archive