Future-H2O at GeoKarlsruhe2021
At the GeoKarlsruhe2021 conference, Annika presented preliminary results of experiments with the Deep Learning model in the "Young Scientist" session. In the experiments, different data combinations and hyperparameters of the Long-Short-Term-Memory (LSTM) network model were tested with respect to model performance in the task of groundwater level (GWL) forecasting. Data combinations included up to 1200 GWL time series from coastal wells as well as input data of environmental attributes at the selected well sites. Preliminary results of the experiments are that the LSTM model is able to generalize groundwater recharge and discharge processes when it is trained on multiple well sites simultaneously. Further experiments are needed to find out to what extent generalization of natural processes and local anthropogenic impacts is (not) possible. Predictions of GWLs of at least moderate quality were obtained for most wells considered to be unaffected by pumping activities (Figure 1).
Figure 1 Examples of observed and simulated groundwater level (GWL - in meters below natural surface) dynamics (left) and Nash-Sutcliffe efficiency (right) of simulations at natural well sites with a deep-learning model based on data from 1200 wells (red) and a deep-learning model based on only 79 natural wells (green). © Nolte, A.