Intelligent Systems

Project: Prediction of water levels using examples from the region of "Willenscharen"



  • Funded by: State of Schleswig-Holstein

  • Partner: Landesamt für Landwirtschaft, Umwelt und ländliche Räume (LLUR)

  • 1st Period: October 2020 - February 2021, 2nd Period: March - September 2021

  • Responsible: Michel Spils, Simon Reichhuber

  • 1st Period: Final report

  • 2nd Period: Final report


Schleswig-Holstein is a federal-state determined by water, with about 32,000 km of running water, more than 1400 km of coastline, about 300 lakes and about 25% of flood-prone land area. Most areas are considered small and difficult to predict with classical models. Also, these models are expensive to operate and maintain.

Within the scope of this project, it shall be investigated to what extent a short-term prediction for hydrological gauges/levels based on machine learning methods can be achieved. Artificial neural networks offer a particularly promising possibility. While other methods are based on physical principles and try to map them mathematically, neural networks are independent of physical relationships. In small areas surrounding the level and partly due to backwater from the many lakes or coastal waters, the underlying physical assumptions are often not continuously given.

Therefore, the basic hypothesis of this project is that a self-learning level prediction is possible with the inclusion of hydrological knowledge and input data using an artificial neural network. A literature search and first LLUR experiments show the general suitability of the method. However, the question remains whether an artificial neural network can be built up for water levels at gauges with their own distribution of data and the demand for the most precise prediction of flood peaks in the typical catchment areas of Schleswig-Holstein, which meets the requirements of hydrological predictions.

The prediction of the level Willenscharen is considered as an example. The main focus of the research work is on the most precise representation of flood events about the height of the flood peak and the temporal correspondence between the model and the measured data. The prediction quality shall be optimized. This is to be achieved by determining and using a suitable neural network and by selecting and determining suitable input parameters. Due to the time factor in flood formation, the input parameters must be considered in their temporal course.