Short-term precipitation forecasting with deep neural nets

Forecasts of precipitation the next minutes and hours ahead are of great value for a wide range of users including the public. At the Norwegian Meteorological Institute such forecasts have been generated for several years using optical flow methods and made available in the yr app and at yr.no. In recent years approaches based on deep learning have also demonstrated promising results. For example, the use of convolutional LSTMs has been proposed by Shi et. al. (2015), while Sønderby et. al. (2020) apply networks with attention-based layers.
Master

In this project, the overall objective will be to study, develop and test various deep neural networks on a gridded dataset of radar reflectivities with high temporal resolution. The work will naturally also include analysis of forecast skill and properties. Case studies of extreme precipitation events are also of special interest. The project can to some extent also be adapted to the student's interest.

Goal

The goal is to study, develop, and test various deep neural networks on a gridded dataset of radar reflectivities with high temporal resolution.

Learning outcome

Deep understanding of deep neural networks
Working on a real-world application
Collaboration with researchers
Possibility to implement and research a novel approach

Qualifications

Python programming
Knowledge about deep learning is an advantage

Supervisors

  • Michael Riegler
  • Hugo Hammer
  • John Bjørnar Bremnes, Research scientist, Meteorologisk Institutt, j.b.bremnes@met.no
  • Siri Sofie Eide, Phd candidate, Meteorologisk Institutt, sirise@met.no

Collaboration partners

Meteorologisk Institutt

References

  • Shi, X. et. al. (2015). Convolutional LSTM network: A machine learning approach for precipitation nowcasting. Advances in neural information processing systems 28, 802-810.
  • Sønderby, C. K. et. al. (2020). MetNet: A Neural Weather Model for Precipitation Forecasting, arxiv.org/abs/2003.12140

Contact person