Navigation and service


Efficient Local Precipitation Prediction through Machine Learning

Due to the stochastic nature of atmospheric circulation, probabilistic precipitation predictions (where, when, how much with which probability) provide significantly more information than classical deterministic predictions, which do not capture any data on the uncertainty of the forecast. However, heavy precipitation events or longer periods of non-rainfall can have enormous economic consequences, so that the assessment of uncertainty is of decisive importance. DeepRain will combine modern methods of machine learning with high-performance data provisioning and processing systems to generate spatially and temporally high-resolution maps with improved and validated precipitation predictions including their uncertainties based on high-resolution regional weather models. In addition to the actual method development, aspects of data curation and efficiency will be specifically investigated in order to demonstrate a complete process chain at the end of the project, which can be transferred into operational use or embedded in existing workflows. The basis for the application of machine learning is a novel combination of data from numerical weather forecast models with precipitation radar, lightning and station measurements as well as topographic data.

The concrete goal of the project is the calculation of probabilities for strong, medium and weak rainfalls (including the expected precipitation) and for the occurrence of thunderstorms in the whole federal territory over a period of 24 hours with a spatial resolution of 1 km or better. Further focal points are the evaluation of the system and the performance optimization of the data preparation and the machine learning algorithms. The latter is essential, since only the performance of current HPC architectures makes such an approach possible. However, the required performance can only be achieved if the programs and algorithms are adapted to the architectures.

- Forschungszentrum Jülich GmbH, Germany
The contact person is Martin Schultz

- Deutscher Wetterdienst, Hans-Ertel-Zentrum für Wetterforschung
- Universität Osnabrück, Institute of Cognitive Science
- Jacobs Universität Bremen, Large Scale Scientific Information Systems
- Universität Bonn, Meteorologisches Institut

DeepRain is funded by the Bundesministerium fuer Bildung und Forschung (BMBF) under grant agreement 01 IS18047A-E.

The grant period is October 2018 until September 2021.

More detailed information about the project is available at the project's homepage.