Comparative applications of data-driven models representing water table fluctuations

In this study, several advanced data-driven models are adopted for estimating water table levels in conjunction with a proposed data preprocessing procedure that includes detrending, deseasonalization, normalization, outlier exclusion, and formatting. To verify the proposed strategy, a non-linear auto-regressive exogenous model based on deep neural networks (NARX-DNNs), a long-short term memory (LSTM) model, a gated recurrent unit (GRU) model, and a reference model based on an auto-regressive exogenous (ARX) model are comparatively applied to water table level time series from the Jindo Uisin and Pohang Gibuk monitoring wells (years 2005–2014). To test the developed preprocessing method, estimates with and without the proposed detrending and deseasonalization (DTDS) are compared quantitatively. In the comparative applications, all four models show reasonable prediction accuracies. In addition, it is found that the estimations from the NARX and LSTM models are superior to those of the other models in terms of prediction accuracy, regardless of whether DTDS is adopted. In both data sets, there is water table level depression during 2014 due to drought throughout the entire Korean peninsula. In multiple analyses, stressed aquifer conditions are identified through estimations based on differences between estimates and observations, where the differences are found to be more obvious with DTDS preprocessing. Thus, by using the proposed preprocessing method, hydrologically stressed conditions in an aquifer can be effectively noticed at an earlier stage. The results show that the advanced data-driven models can be more effective when adopted in conjunction with the proposed preprocessing method and successfully utilized for monitoring and management of groundwater resources.