SNIC
SUPR
SNIC SUPR
Machine learning for weather forecasting and classification
Dnr:

SNIC 2018/5-99

Type:

SNAC Small

Principal Investigator:

Sebastian Scher

Affiliation:

Stockholms universitet

Start Date:

2018-08-01

End Date:

2019-02-01

Primary Classification:

10508: Meteorology and Atmospheric Sciences

Webpage:

Allocation

Abstract

For a small research project, that I had carried out in the last half year, I would like to apply for additional computation time in order to finish a revision of a submitted paper. In order to complete the revision and to try out some related things, I would like to apply for 6 more months. I need only relatively little computation time, but potentially try different things spread out over a couple of months. Project abstract: The quality of weather forecasts not only depends on the quality of the initial conditions and the accuracy of the model, but also on the current state of the atmosphere. Thus, the uncertainty of weather forecasts is correlated to the meteorological situation in which the forecast is started. This relationship is, however, highly nonlinear. Due to the large amount of weather forecasts that are available from the past, it should be possible to detect certain patterns in the initial conditions of the forecast that lead to especially good or bad forecasts. (Deep) artificial neural networks have been very successful in many machine learning tasks, especially image recognition, and are applicable to highly non-linear problems. This, and the fact that atmospheric states can be represented on grids, which can be interpreted as images, makes it very interesting to use neural networks, especially convolution networks, for finding relations between atmospheric states and forecast quality. The goal is to train a neural network on past predictions, and then use the network to predict the forecast error of operational weather forecasts.