University of Twente Student Theses
SG-AIFS : A High-Resolution Deep Learning Approach to Weather Forecasting in Western Europe.
Buurman, S. (2024) SG-AIFS : A High-Resolution Deep Learning Approach to Weather Forecasting in Western Europe.
|
PDF
12MB |
| Abstract: | Traditionally, weather forecasting relies on numerical weather prediction (NWP) models. However, recent advances in the field have demonstrated that deep-learning based weather prediction (DLWP) models can successfully be trained on historical re-analysis (ERA5, 0.25 degree resolution) data, to produce accurate global medium-range forecasts. Two DLWP approaches can be distinguished: Graph Neural Networks (GNNs) and Transformer models. Expanding on these approaches, the European Centre for Medium-range Weather Forecasts (ECMWF) has developed a global Graph-Transformer model based on GraphCast from Google DeepMind, called the Artificial Intelligence Forecasting System (AIFS). This MSc thesis investigates the possibility of extending these new developments to high-resolution modeling on a limited domain, by adapting AIFS to include a stretched grid (SG-AIFS) using refined hidden grid layers in the processor step. This research was done in collaboration with MET Norway, whose evaluation on surface observations has outperformed their operational HARMONIE-AROME NWP model on certain variables, although showing underestimation of extremes [29]. In this research, the DOWA (Dutch Offshore Wind Atlas) dataset - a reanalysis from the 2.5-km HARMONIEAROME NWP model of KNMI - is used to integrate with the lower-resolution ERA5 data and is subsequently connected to the stretched grid. First, different processor refinements are assessed by evaluating hidden grid sizes using ERA5 data. We find that although increasing processor refinements accelerates training time, it results in marginal improvements over longer lead times. On the other hand, rollout training proved essential in reducing RMSE values across all lead times. These results are employed to train the model on the DOWA dataset, producing high-resolution deterministic forecasts whilst minimizing computational resources. The model provides +6h predictions with some accuracy, although it lacks detailed features and longer lead times show artefacts resembling the processor hidden grid structure. |
| Item Type: | Essay (Master) |
| Clients: | KNMI, De Bilt, Netherlands |
| Faculty: | EEMCS: Electrical Engineering, Mathematics and Computer Science |
| Subject: | 31 mathematics, 38 earth sciences, 54 computer science |
| Programme: | Applied Mathematics MSc (60348) |
| Link to this item: | https://purl.utwente.nl/essays/104770 |
| Export this item as: | BibTeX EndNote HTML Citation Reference Manager |
Show download statistics for this publication
Show download statistics for this publicationRepository Staff Only: item control page