Master thesis defense by Wiebke Margitta Kolbe
Title: Testing the usage of neural networks in the shortwave radiation parameterization of the WRF model
Abstract: Radiative transfers in the atmosphere are difficult to compute accurately in numerical weather prediction (NWP) models, without the procedure becoming too computationally expensive.
In this thesis it has therefore been tested to substitute a part of the shortwave radiation parameterization in the Weather Research and Forecasting (WRF) model with neural networks, to investigate a possible increase in computational efficiency of such a modified parameterization and its accuracy.
The data set used to train the neural networks was created with the RRTMG-fast shortwave radiation parameterization scheme in the WRF model.
After several optimization processes, three configurations of neural networks were implemented and tested in the WRF model, replacing an computationally expensive part of the RRTMG-fast scheme.
To evaluate the three neural network modified shortwave schemes, four 96-hour simulations were carried out as case studies, to compare how the model performs in different weather situations.
Additionally to the original RRTMG-fast scheme and the three variants modified with neural networks, the four case studies were simulated with three other shortwave parameterization schemes as well: the RRTMG, New Goddard and Dudhia schemes.
Comparison of the results showed that the modified neural network schemes were able to make predictions similar to the original RRTMG-fast scheme, but were computationally slower.
A quick test addressed one of the causes, the activation function, and suggested that the computational time of the neural networks can be reduced significantly by using a different activation, though the new performance has yet to be evaluated, while possible further optimizations are addressed.
Supervisor
Eigil Kaas
Censor
Peter Aakjær
Participating from room 008 or via Zoom by using the following link: https://ucph-ku.zoom.us/j/64628535009