Back to the programme printer.gif Print




Delegates are invited to meet and discuss with the poster presenters in this topic directly after the session 'The model chain: First steps towards tomorrow's technology' taking place on Thursday, 13 March 2014 at 09:00-10:30. The meet-the-authors will take place in the poster area.

Jeremy Sack EWC Weather Consult GmbH, Germany
Co-authors:
Jeremy Sack (1) F P Achim Strunk (1) Jon Meis (1) Anton Kaifel (1)
(1) EWC Weather Consult GmbH, Karlsruhe, Germany (2) Centre for Solar Energy and Hydrogen Research (ZSW), Stuttgart, Germany

Printer friendly version: printer.gif Print

Abstract

From uncertainties in the model chain to probabilistic forecasts

Introduction

Probabilistic forecasting has become state-of-the-art in wind energy prediction. Several studies have shown that probabilistic forecasts do not only provide more extensive information about future weather development, but also allow players on the wind energy market such as grid operators, risk managers, energy traders and others to save valuable resources.


Approach

The use of ensemble Numerical Weather Predictions (NWP) has turned out to be an elegant way of representing the forecast uncertainties depending on prevailing weather situations. It is now common practice to derive predictive densities from the ensemble's set of trajectories and to maximize the skill through a calibration with measurements. Although expensive in terms of computing resources and money, ensembles including a large set of members have been preferred in recent studies.

Main body of abstract

The present study introduces a way of generating probabilistic forecasts obtained from a small multi-model-ensemble offering similar skill as a large single-model ensemble (ECMWF_EPS). This is achieved by increasing the number of ensemble members through varying more parts of the model chain than only the NWP forecasts. Although the latter account for the largest portion of the forecast errors, it turns out that variations in downscaling, interpolation and the conversion of model parameters to power yield valuable additional probabilistic information. As a next step, the increased set of ensemble members is optimized through an adaptive calibration of (u,v)-wind ensemble forecasts (Pinson, 2012) followed by a conditional weighted combination of wind power forecasts (Thordason, et al. 2008). The latter makes use of each member of the multi-model ensemble having distinct characteristics. This allows to associate different uncertainties and weights with each member, which drastically improves the skill and which is not possible for ECMWF_EPS.

Conclusion

Depending on lead time and location, this low-cost method yields comparable or even better results than those originating from ECMWF_EPS with respect to the continuous rank probability skill score. The question of how crucial expensive large single-model ensembles are in an operational wind power forecasting environment is discussed and a method for choosing the best option is proposed.


Learning objectives
The present study offers insights into the different uncertainties in a classic wind power forecasting setup, into the problem of associating uncertainties to the non-linear and bounded variable wind power, into the calibration of ensembles and into the evaluation of probabilistic forecast information.