Skip to main content

Machine Learning (LTM) and rainfall runoff modelling

Are you interested in this MSc topic? Send an email with your CV to Jaap.Kwadijk@deltares.nlm.j.booij@utwente.nl, ruben.dahm@deltares.nl 

Participating institutes: Deltares, University of Twente

Areas under investigation: the Rhine basin and other basins

Rationale and objective

Hydrological modeling has long been about integrating accumulated knowledge of water stores and physical processes into schematizations. Recently, research done by Kratzert et al. [1] has shown that an LSTM neural network that had no such encoded knowledge, was able to outperform classical hydrological models. By training on the publicly available CAMELS dataset [2], the model was able to learn hydrological relationships from a diverse set of catchments. These results have been encouraging enough for the authors to opine about “What Role Does Hydrological Science Play in the Age of Machine Learning?” [3].

Last year, we have started exploring these methods and carried out initial comparisons with simulation results from the WFLOW distributed hydrological model, which showed very promising results. In this thesis we want to address questions related to the applicability of these methods in practice, such that we will know when to use it compared to traditional models, what the limitations are, and how we can apply it effectively outside of the areas of the existing training datasets, such as the Rhine basin.

Approach

The student will start reviewing the literature available on the subject. To be able to apply this method in a wide range of areas, python scripts will be developed to prepare suitable training datasets. The model performance will be compared to traditional models as well as observations, focusing both on high and low flows.

Deliverables

The student will deliver a thesis that answers the research questions, as well as the scripts and datasets that were used to get these answers, such that we can reproduce the results.

Requirements

  • Successfully completed courses on hydrology and/or hydrological modellling and/or water resources management
  • A reasonable acquaintance with Python or other script languages for programming

References

[1] Kratzert et al. 2019 - Towards learning universal, regional, and local hydrological behaviors via machine learning applied to large-sample datasets

[2] CAMELS: catchment attributes and meteorology for large-sample studies

[3] Nearing et al. 2020 - What Role Does Hydrological Science Play in the Age of Machine Learning?