Development and application of machine learning based algorithms for the prediction of groundwater levels

Project description

The reliability of modelling and prediction of groundwater levels with physical-numerical groundwater flow models is strongly dependent on the availability of field data for model parameterisation. Especially in studies on regional and supra-regional scales, models often fail due to missing or incomplete information regarding the spatial extent of the individual geological formations and their geohydraulic properties. Approaches based on artificial intelligence, such as artificial neural networks (KNN), are an alternative here, as they reduce the question to an input-output relationship, while descriptions of the physical process are omitted. KNN are able to learn and map linear and nonlinear relationships in complex systems. A reliable prediction of groundwater levels is the basis for e.g. the derivation of water availability for drinking water supply and irrigation requirements for agriculture, the delineation of potential land settlement zones by extremely low groundwater levels in connection with droughts and/or pumps, the delineation of areas of potential groundwater peaks for transport infrastructure, buildings and agricultural land, as well as the development of suitable avoidance and adaptation strategies.

The aim of the project is to develop a method that allows punctual forecasts of groundwater levels in Germany for medium scales and area-based interpretations at small scales (1:1 million, 1:1.5 million). Based on the individually available groundwater dynamics, the time series of all available monitoring stations are first grouped using a machine learning based methodology and then representative monitoring stations are selected. These so-called reference monitoring sites (RM) reflect the groundwater dynamics dominating in the respective group and are predicted representative for the group. The KNN-based RM prediction is then transferred to other group members. The actual predictions cover time spans from one week up to three months (seasonal predictions). Data basis are mainly groundwater data of the national monitoring networks and meteorological measurement and modelling data of the German Weather Service (DWD).

Machine-learning based grouping of groundwater level time series according to dynamics (top) and also machine-learning based prediction of groundwater levels (bottom)