Using neural networks with latent variables to predict future radar frames
Abstract
The United Kingdom has a very good radar network coverage, with a continually growing amount of available precipitation observation data. Machine learning offers an opportunity to harness more of the potential of this valuable data, particularly for nowcasting applications. The aim of this work is to understand if and how deep learning methods could add value over more traditional nowcasting methods. Work-to-date has focussed on using UK radar composite data, trying to predict future radar frames from an input of previous timesteps of radar data.
Results from a complex neural network that makes use of latent space to represent the uncertainty in the system will be presented. Early verification results and example case studies will be shown, comparing the machine learning model predictions to the Met Office's current operational precipitation nowcast. An ensemble approach using this model will also be presented as well as consideration of how to handle the higher precipitation rates that can lead to higher impact events.- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2020
- Bibcode:
- 2020AGUFMH147...08B
- Keywords:
-
- 3305 Climate change and variability;
- ATMOSPHERIC PROCESSES;
- 1817 Extreme events;
- HYDROLOGY;
- 1854 Precipitation;
- HYDROLOGY;
- 4318 Statistical analysis;
- NATURAL HAZARDS