Project: Extrapolation in Machine Learning-Based Weather Prediction Models
Supervisors: Nicola Gnecco (Imperial) and Almut Veraart (Imperial)
Project Description:
Recently, there have been significant advances in machine learning-based weather prediction (MLWP) methods (e.g., Keisler, R., 2022; Lam, R. et al., 2022; and Lang, S. et al., 2024).
These approaches train graph neural networks (GNN) globally at a high spatial resolution using several years of historical data from the ECMWF’s ERA5 reanalysis database. One key advantage of MLWP is the ability to generate 10-day forecasts that are as accurate as state-of-the-art numerical weather prediction methods but in a fraction of the time. Despite their success, MLWP methods struggle to predict extreme weather events that have never appeared in the training data.
Our research question focuses on developing MLWP techniques capable of accurately predicting these extreme events by extrapolating beyond the historical data range. To reach this goal, we will model specific regions of interest (Oskarsson, J. et al., 2023) to keep the computational demands feasible and build on the open-source project (https://github.com/mllam/neural-lam).
Project envision three research directions.
- Computational: Adapting the ensemble boosting method proposed by Fischer, E. et al. (2023) to enhance historical datasets with extreme observations and train a GNN for regional predictions.
- Methodological: Extending the Engression methodology by Shen, X., & Meinshausen, N. (2023) to work with GNN and apply them to weather forecasting in specific regions.
- Theoretical: Adapting the Progression methodology by Buriticá, G., & Engelke, S. (2024) to work with non-parametric models, potentially through linear representation in kernel spaces.
