Year: 2024
Type: dissertation
Abstract: This thesis explores the application of generative modeling techniques to probabilistic weather forecasting, specifically through the development of DEFfusion (Direct Ensemble Forecasting with Diffusion). DEFfusion aims to address limitations in current Machine Learning Weather Prediction (MLWP) models, which tend to produce blurry forecasts and rely heavily on iterative methods. By using a diffusion model, the thesis demonstrates the potential of direct forecasting to generate more accurate and computationally efficient weather predictions. The approach involves generating an ensemble of possible future weather states from a given initial condition, utilizing a Quasi-Geostrophic (QG) model for experimental feasibility. The methodology includes training a neural network to perform dimensionality reduction via an autoencoder and employing a diffusion model to handle the probabilistic aspects of forecasting. Evaluation metrics such as Root Mean Squared Error (RMSE), Continuous Ranked Probability Score (CRPS), and Skill/Spread ratio indicate that DEFfusion offers improved performance over iterative approaches. This research contributes to advancing MLWP models by enhancing their ability to quantify uncertainty and capture extreme events, paving the way for more reliable and efficient weather prediction systems.
Author: Andrae, Martin
Language: English
Subfield: Atmospheric Science
Field: Earth and Planetary Sciences
Domain: Physical Sciences
SDG: Climate action
Open Access status: green
OpenAlex ID: W7072214465