General summary
In order to achieve optimum operation for wind turbine generators the suitability and cost of key components should be quantified enabling fair comparison of decision options. For WTGs operating in cold climates the skill in detecting adverse icing conditions at an early enough stage is especially important when e.g. controlling the risk of ice throw or when avoiding prolonged production losses. In this presentation we attempt to quantify the uncertainty and evaluate the performance of an icing forecast system. Simulated icing have been translated into a traffic light alarm system for 20+ wind farm sites. More than 300 individual wind turbines in Sweden, Norway and Finland have been analyzed.
Pro-active uses for a proper detection system(s) include risk mitigation such as the operational strategy of stopping the turbines in time before icing occurs, in turbine controllers for preemptive heating of turbine blades, or simply to trigger closer monitoring and awareness depending on the forecasted icing severity.
Method
Hindcast data is produced for the period 2000-present by the use of the Weather Research and Forecasting model (WRF). The grid resolution of the forecasts is 4 km x 4km. From these hindcast dataset, icing have been calculated as the icing rate on a freely rotating cylinder according to ISO 12494 – Atmospheric Icing on Structures. A local adjustment of the atmospheric moisture content and liquid water content has been carried out to adjust for the subgrid scale terrain in the model and thus make the icing forecasts representative for typical wind power sites also in complex terrain.
Results
The skill scores of the icing forecasts are calculated for each of the wind farms and presented in ROC diagrams. The results show that for a majority of the wind farms, 70-80% of all icing periods identified from the SCADA data are also captured by the forecast model. The false alarm rate is typically in the range 5-10 %.
However, the skill scores are sensitive to the threshold settings of the traffic light conditions. By including icing signals from higher model levels or neighboring model grids the hit rates of the icing forecasts are generally improved, but with increase of the false alarm rate.
Conclusions
The skill score of the modeled icing depends on thresholds and the setup of the individual wind farm and can thus be tailored toward the user needs. Proper quantification of uncertainty measures and performance indicators enables fair comparison between options as well as informed decisions and judgments on suitable operational strategies.
The actual icing levels for defining the probabilistic take-action or traffic-light signals will vary from case to case and on the cost-effectiveness of different systems and the choices presented here are therefore considered an exemplification. For example, to handle possible risks of ice throw from turbines it will be important to capture as many as possible of the icing events while the forecasting of icing losses for energy traders will achieve the best value as best possible compromise between the false alarm rates and hit rates.