In this work, we develop a probabilistic system for our short-range forecasts of icing on wind turbines and related wind power production losses. The modelling chain consists of several model components, starting with the numerical weather prediction model HARMONIE-AROME, a physical icing model based on the so-called Makkonen model, and an empirical production model. Forecasts are uncertain owing to errors over the entire modelling chain, especially due to the meteorological initial as well as boundary conditions and model formulations in model components. Probabilistic forecasting provides the statistically best forecast and its uncertainty, and therefore, it is valuable when using short-range forecasts of wind power production in cold climates.
In the first step, we introduced a high-resolution NWP ensemble prediction system into this modelling chain and, thus, included uncertainties from initial conditions and representation applying a neighbourhood method. The ensemble prediction system consists of 11 members and has been run for up to +42 hours for a two-week period in the winter 2011/2012 with a horizontal resolution of 2.5 km over a Swedish domain of 1100x1600 km2. For this period also a neighbourhood method was tested to account for representativeness errors. When validated against wind power production data from three sites, we found that the best forecast skill and forecast uncertainty was provided when both NWP ensemble and neighbourhood method was combined. Here it was especially important to run the icing-production model for each ensemble forecast and gridpoint of the neighbourhood separately.
In the second step, we have employed an uncertainty quantification method called deterministic sampling in order to capture the uncertainty of the icing model. In a literature study, we have identified 5 uncertain parameters of the icing model with an estimate of their mean and standard deviation. Then, we construct an ensemble with deterministic sampling consisting of 9 members that exactly describes the estimated statistical moments of the uncertain parameters. For two other winter periods, 10 weeks 2014-2014 and 12 weeks 2014-2015, we compared this icing model ensemble with deterministic sampling with a random sampling ensemble with 10,000 members. The two ensembles deliver comparable results, but with a fraction of computational costs for the deterministic sampling compared to the random sampling. This becomes especially important, when the icing model ensemble is combined with the NWP ensemble leading to a multiplication of ensemble sizes.
The three different probabilistic forecasting methods have the benefit of both uncertainty estimations for each forecast and higher skill for the ensemble mean compared to a deterministic forecast. Combining different methods accounting for uncertainties in different parts of the modelling chain resulted in even better forecast skill and forecast uncertainty estimations.