Objective & Background
Accurate multi-site wind speed and aggregated wind power forecasts are important for various stakeholders in the energy sector such as grid operators and energy traders. When using purely deterministic forecasts, it must be taken into account that the resulting forecast errors for different locations are spatially correlated and that this effect is most pronounced for long forecast horizons. Probabilistic approaches based on meteorological ensemble forecasts allow considering such spatial dependence structures in a more adequate way. In our analysis, we investigate whether statistical postprocessing methods are able to improve the skill of probabilistic multi-site forecasts compared to the raw ensemble forecast for different strengths of spatial correlation. Furthermore, we analyze the sensitivity of different multivariate scoring rules with respect to varying spatial correlations. For the analysis, we utilize ensemble wind speed forecasts (100m) provided by the European Centre for Medium-Range Weather Forecasts (ECMWF) and 100m wind speed measurements for two German offshore research platforms (FINO2 and FINO3) from February 2010 to July 2014.
Method
We apply the truncated normal Ensemble Model Output Statistics (EMOS) model proposed by Thorarinsdottir and Gneiting (2010) to postprocess the raw ensemble wind speed forecasts separately for each site. In order to recover the information about the spatial dependence structure from the raw ensemble, we further utilize a quantile-based Ensemble Copula Coupling (ECCQ) technique (see Schefzik et al. (2013)). We compare the skill of the different forecasts (RawEnsemble, EMOS, ECCQ) in terms of the Energy Skill Score (ESS) and the Variogram Skill Score (VSS) for forecast horizons from 3 to 120 hours.
Principal Findings
We find that the EMOS forecasts outperform the raw ensemble with respect to the Energy Skill Score for all forecast horizons. For longer forecast horizons, the VSS, however, indicates that the raw ensemble yields better forecasts than EMOS. When applying the ECCQ technique to recover the information about the spatial dependence structure, the forecast skill improves especially for longer forecast horizons compared to EMOS. While the VSS indicates that the quality of the forecasts improves for forecast horizons longer than 50 hours, ECCQ outperforms EMOS only for very long forecast horizons close to 120 hours regarding the ESS. In a sensitivity study, we show that this finding is a direct result from different characteristics of the two scoring rules.
Conclusion
The analysis shows that copula coupling techniques can help to improve the skill of probabilistic multi-site forecasts especially for longer forecast horizons when spatial correlation structures are most pronounced. The results further underline the need for different scoring rules to evaluate multivariate probabilistic forecasts. Due to different characteristics of the Variogram Score and the Energy Score, applicants should interpret and compare these scores in empirical studies with caution. In future research, the findings from our analysis shall be generalized for multivariate settings with more than two locations.