Two Data Science students from the Department of Mathematics presented their BSc project at EMS2022
Thomas Lykke Rasmussen and Mads Emil Marker Jungersen study Data Science at the Department of Mathematics. On 4-9 September they will both participate in the congress EMS2022 (European Meteorological Society) to be held on the Poppelsdorf Campus of the University of Bonn in Germany. Here, they have just presented their bachelor project "Cloud Cover Nowcasting with Deep Learning", which has attracted much interest.
"It is an impressive project that Thomas and Mads Emil have made. It is very unusual that a BSc project has such a high news value and is of such a quality that it can provide the basis of a paper chosen for an international research conference in free competition with researchers from all over the world."
– Henrik Karstoft, supervisor.
The project studies how computer vision and deep learning can be used to predict cloud cover, using satellite images. Being able to predict cloud cover from satellite images has received increased interest in the transition to sustainable forms of energy such as solar energy. Although deep learning has been successfully applied in several meteorological tasks on various occasions, few studies have investigated the problem of cloud cover prediction based on satellite data. In this project, Thomas and Mads Emil have investigated how different deep learning architectures, such as ConvLSTM, U-Net and MetNet, are used to predict cloud cover over Germany over a time frame of 90 minutes. In addition, the project investigates minor changes in the deep learning model ConLSTM. Thomas and Mads Emil have, for instance, investigated how many frames from the past and how large a section of Europe's cloud covers, the model requires to be able to predict Germany's cloud cover. They have also investigated the effect of adding various information to the model about the area it predicts, e.g., where there is water or mountains. Finally, they have also investigated the effect of varying the loss function, which is the cost function you try to minimize when training a neural network. Thomas and Mads Emil tested all their models with known weather models and have thereby formed a basis for the direction of future research in the prediction of cloud cover.
"It has been exciting to use the complex deep learning models to predict something as familiar as the weather."
– Thomas Lykke Rasmussen
"Working with the project, which has involved various aspects of data collection, data pre-processing, analysis, and interpretation, has been a truly educational process."
– Mads Emil Marker Jungersen