Machine Learning-Based Reconstruction and Prediction of Groundwater Time Series in the Allertal, Germany

Abstract

Listen

Translate article

State-of-the-art hydrogeological investigations use transient calibrated numerical flow and transport models for multiple scenario analyses. However, the transient calibration of numerical flow and transport models still requires consistent long-term groundwater time series, which are often not available or contain data gaps, thus reducing the robustness and confidence of the numerical model. This study presents a data-driven approach for the reconstruction and prediction of gaps in a discontinuous groundwater level time series at a monitoring station in the Allertal (Saxony-Anhalt, Germany). Deep Learning and classical machine learning (ML) approaches (artificial neural networks (TensorFlow, PyTorch), the ensemble method (Random Forest), boosting method (eXtreme gradient boosting (XGBoost)), and Multiple Linear Regression) are used. Precipitation and groundwater level time series from two neighboring monitoring stations serve as input data for the prediction and reconstruction. A comparative analysis shows that the input data from one measuring station enable the reconstruction and prediction of the missing groundwater levels with good to satisfactory accuracy. Due to a higher correlation between this station and the station to be predicted, its input data lead to better adapted models than those of the second station. If the time series of the second station are used as model inputs, the results show slightly lower correlations for training, testing and, prediction. All machine learning models show a similar qualitative behavior with lower fluctuations during the hydrological summer months. The successfully reconstructed and predicted time series can be used for transient calibration of numerical flow and transport models in the Allertal (e.g., for the overlying rocks of the Morsleben Nuclear Waste Repository). This could lead to greater acceptance, reliability, and confidence in further numerical studies, potentially addressing the influence of the overburden acting as a barrier to radioactive substances.