Mitigating climate risks by improving weather forecasts using copulabased approaches for post-processing (PP) of forecast ensembles

点击翻译按钮获取中文摘要

Accurate weather predictions play an important role in understanding and mitigating risks induced by climate change, as well as for predicting power outputs of renewable energies. Weather prediction today is conducted via numerical weather prediction (NWP) models. The output of a model run is a single deterministic forecast of future weather events or variables. To be able to assess forecast uncertainty it has become common practise to use an ensemble of NWP forecasts obtained by running the NWP model multiple times with different initial conditions and/or model formulations. However, these so-called ensemble forecasts typically lack calibration and require statistical postprocessing. Several statistical postprocessing methods have been developed to account for the needs of different weather variables. It has become specifically important to extend the postprocessing models so that they explicitly incorporate dependencies e.g. in space, time or between weather variables. This project aims at developing new types of postprocessing models based on vine copulas. Vine-copulas allow for very flexible and data driven modelling of any type of multivariate dependence. The aim is to adapt the vine copulas for use in the context of statistical ensemble postprocessing. Specifically, it is planned to develop vine copula based postprocessing models for different weather variables, such as temperature, wind speed, precipitation, cloud cover and solar irradiance. The vine copula based quantile regression will also be utilized for postprocessing renewable energy weather variables like wind speed and solar irradiance and performing conversion to prediction of the respective power output in a joint approach. In a next step these models are supposed to be extended to the multivariate context, incorporating dependencies in time, in space, and between weather variables and also attempting to model all these dependencies jointly. The developed models are supposed to be implemented within the statistics software package R, and to be compared to state-of-the-art postprocessing models in a study investigating predictive performance a calibration properties of the postprocessed forecasts.