Probability forecasts of ice accretion on wind turbines derived from multiphysics and neighbourhood ensembles

Abstract

This study explores the potential of a multiphysics regional ensemble prediction system to improve forecasts of wind turbine icing, examining several error-representation schemes to capture the forecasting uncertainties of the icing process. An 11-member multiphysics ensemble based on the Weather Research and Forecasting (WRF) model is run for two winter periods over Europe. Regional verification of surface variables shows that parametrization diversity makes the multiphysics ensemble less underdispersive compared with the European Centre for Medium-Range Weather Forecasts (ECMWF) global ensemble, without deteriorating the overall forecast accuracy significantly, in particular at forecast ranges below 36 hr. Probability forecasts of active ice growth in the day-ahead time range (12–36 hr) are derived for two wind farms on hilly terrain in Central Europe and their skill is assessed in terms of relative operating characteristics, reliability, and potential economic value (PEV). Probability forecasts enhance the maximum PEV significantly, but the improvement of the multiphysics ensemble seems modest compared with a simple neighbourhood ensemble approach. Icing forecasts are affected by a considerable degree of overconfidence, meaning that forecast probabilities cannot be used at face value, but require calibration for users to draw benefit from them. The multiphysics ensemble fares slightly better in this regard; however, results point to persistent ensemble underdispersiveness and yet underrepresented forecast uncertainty. Overall, findings show that a large portion of the gain in skill through the use of probabilistic icing forecasts is obtained with a computationally cheap neighbourhood method, a technique easily accessible to forecast users without complex ensemble prediction systems.