Predicting groundwater levels using linear regression and neural networks

Water resources managers can benefit from accurate prediction of the availability of groundwater. In this project I present two models to predict groundwater levels in an unconfined shallow aquifer in the Searsville basin, part of the Jasper Ridge Biological Preserve. The input data (ie, features) for the models includes local weather, lake stage, and stream flow data, and moving averages of the weather, stage and stream flow data taken over time-frames of one week, one month, three months and six months. When moving averages are included as features, a linear regression model does well at predicting summer groundwater levels. In contrast, a feed-forward time-delay neural network does well at predicting winter groundwater levels. In combination, these models can provide useful predictions for groundwater levels throughout the year. Feature analysis indicates that the most important features are the longer time-frame moving averages that measure the “seasonality” of the example.