Understanding and characterizing groundwater system properties is of great importance to develop sustainable groundwater management strategies. For this purpose, groundwater hydrographs are a valuable source of knowledge, since they contain information about system properties (e.g. aquifer type), artificial (e.g. withdrawal/infiltration) and natural environmental factors (e.g. groundwater-streamflow interaction). Such factors interact and superimpose temporally and spatially, which makes determining the individual contributions a challenging task. However, understanding spatial dynamics patterns is a precious source of information for this purpose. Generally, in many regions, large amounts of groundwater data with high resolution in time and space are available but lack an adequate set of tools for analysis. Data driven models are possibly suited to fill in this gap. We developed a machine learning based ensemble-modelling approach to characterize and cluster groundwater hydrographs on regional scale according to their dynamics. We apply feature-based clustering to reduce data quality requirements and to improve exploitation of heterogeneous datasets. ... mehrSuch features describe hydrograph dynamics and serve as surrogates for clustering based on Self-Organizing-Maps and DS2L-SOM-enrichment for cluster determination. Ensemble modeling assures highly robust cluster results, even for real world observational networks undergoing changes. The test area of the method is the Upper Rhine Graben in Germany/France, using more than 1800 weekly sampled hydrographs in the period of 1986 to 2016. The majority shows lengths of almost 30 years, minimum length is six years. Results show that our approach is capable to identify homogeneous groups of hydrograph dynamics. The resulting clusters showed both known and unknown patterns, of which some correspond to certain environmental factors. However, we also discovered new patterns with unknown origin, which need further examination. Possible application of the found groundwater patterns could be for example regional groundwater forecasting by selecting and predicting representative group members. By adapting the describing features, this data-driven method is easily transferrable to other time-series-clustering frameworks.