A0686
Title: D-vine copula based probabilistic weather forecasting
Authors: Annette Moeller - Bielefeld University (Germany) [presenting]
David Jobst - University of Hildesheim (Germany)
Juergen Gross - University of Hildesheim (Germany)
Abstract: Current practice in predicting future weather is the use of numerical weather prediction (NWP) models. These models are run multiple times with different initial conditions and/or model formulations to obtain an ensemble of forecasts that represents model and forecast uncertainty. The resulting ensemble forecasts typically lack calibration and need to be corrected by statistical postprocessing models. A D-vine copula-based quantile regression (DVQR) approach is proposed for ensemble postprocessing. The DVQR incorporates important predictor variables from a large set of potentially relevant ones using a sequential forward selection procedure. It is highly data-driven and allows for the adoption of more general dependence structures as the state-of-the-art ensemble model output statistic (EMOS) postprocessing model. However, the current DVQR does not explicitly allow for the accounting of additional covariate effects, e.g. temporal or spatiotemporal information. Thus, an extension of the DVQR is introduced, where the bivariate copulas are parametrized in the D-vine copula through Kendall's tau, which can then be linked to additional covariates via generalized additive models (GAMs). The performance of the GAM-DVQR is illustrated in a case study for postprocessing of 2m surface temperature forecasts. Different GAM-DVQR models are compared to the benchmark methods EMOS, its gradient-boosted extension (EMOS-GB) and basic DVQR.
