Subproject 2: Generation and preparation of probabilistic weather and power forecasts
The focus of subproject 2 is to generate, evaluate, and calibrate probabilistic products based on weather and power forecasts.
Forecasts inevitably contain uncertainties. These result from uncertainties in the initial and boundary conditions as well as from uncertainties in the model itself coming from, for example, parameterization of sub-scale processes or approximations in the physical equations. Ensemble prediction systems (EPS) are appropriate tools to quantify forecast uncertainty. The different ensemble members can be seen as different deterministic forecasts. A larger spread between the ensemble members indicates that the forecast for this situation is more uncertain. Probabilistic products can be derived from EPS. Typical products that are used to provide a comprehensive description of the forecast statistic are the ensemble mean and variability, exceedance probability based on a defined threshold, quantiles and extreme scenarios. In comparison to the deterministic forecast, such products provide additional information that is required for the integration of renewables into the electricity grid.
In this subproject, the ensemble prediction system COSMO-DE-EPS is optimized to produce probabilistic weather and power forecast products that are relevant for energy applications. COSMO-DE-EPS (see picture below) is a convection-permitting model with a horizontal resolution of 2.8 km. It has been developed at DWD and has been running operationally since 2012. An update cycle of 3 hours is applied and the forecasting horizon is 27 hours (recently changed from 21 hours). This setup is appropriate for the intra-day forecasts, whereas selected runs will be extended to cover the day-ahead.
Since wind and PV-power forecasts do not follow exact theoretical error distributions, the probabilistic power forecasts will be based on quantile regression coupled to artificial intelligence or data-mining methods. Additionally, methods such as kernel density estimation, ensemble dressing, or Bayesian model averaging, which have proved useful in meteorological applications, will be evaluated in this subproject.
Back to "Project"