New publication “Bivariate postprocessing of wind vectors” in the Quaterly Journal of Royal Meteorological Society!
Nowadays, weather forecasts are generated by so-called ensemble prediction systems, which produce a forecast ensemble to quantify uncertainty. However, these forecasts are often biased and underdispersive. Therefore statistical postprocessing methods are applied to improve their calibration and accuracy. Recent research increasingly focuses on multivariate methods that explicitly model dependencies between related variables. In wind forecasting, this is particularly important because the zonal u-component and meridional v-component jointly describe the wind vector, which allows to better capture their dependence structure and therefore produce more realistic forecasts.
Key Contributions
In our new publication, “Bivariate Postprocessing of Wind Vectors”, we introduce and compare three novel approaches for the joint postprocessing of the zonal and meridional wind components. These methods extend the classical Ensemble Model Output Statistics (EMOS) framework beyond standard bivariate formulations that either assume independence between both wind vector components (IND-EMOS) or explicitly account for their dependence structure (BIV-EMOS). Specifically, we propose:
- a gradient-boosted extension of EMOS assuming the bivariate normal distribution (BIV-EMOS-GB),
- a distributional regression network assuming the bivariate normal distribution (BIV-DRN),
- simplified Y-vine copula models (BIV-YV-G, BIV-YV-P, BIV-YV-ALL) to flexibly model dependence between the wind vector components.
In a case study based on 60 observation stations in Germany, we evaluate these methods with respect to calibration, sharpness as well as forecast skill using various (multivariate) verification measures.
Key Findings
- The IND-EMOS and BIV-EMOS, which rely solely on the forecast ensembles of the zonal and meridional wind components as covariates, are consistently outperformed by all alternative approaches that additionally incorporate ensemble forecasts of other meteorological variables as predictors.
- BIV-DRN achieves the strongest overall predictive performance, yielding the best results in terms of the energy score, variogram score, and logarithmic score, see also Figure 1.
- The BIV-EMOS-GB provides the sharpest forecasts given by the lowest determinant sharpness, see also Figure 1.
- The simplified Y-vine copula models (BIV-YV-P and BIV-YV-ALL) provide the best calibrated forecasts, benefiting from the flexibility of selecting non-Gaussian copula families.
