Statistical Postprocessing of Numerical Weather Prediction Forecasts using Machine Learning

Nowadays, weather prediction is based on numerical models of the physics of the atmosphere. These models are usually run multiple times based on randomly perturbed initial conditions. The resulting so-called ensemble forecasts represent distinct scenarios of the future and provide probabilistic projections. However, these forecasts are subject to systematic errors such as biases and they are often unable to quantify the forecast uncertainty adequately. Statistical postprocessing methods aim to exploit structure in past pairs of forecasts and observations to correct these errors when applied to future forecasts.
In this thesis, we develop statistical postprocessing methods based on the central paradigm of probabilistic forecasting, that is, to maximize the sharpness subject to calibration. A wide range of statistical and machine learning methods is presented with a focus on novel neural network-based postprocessing techniques. In particular, we analyze the aggregation of distributional forecasts from neural network ensembles and develop statistical postprocessing methods for ensemble forecasts of wind gusts, with a focus on European winter storms.