Estimates of energy losses due to icing are of major importance in energy calculations for wind parks in Finland. Improved icing loss estimates will result in reduced uncertainty in annual energy production estimates. In this ongoing research project, a modeling tool for icing losses in Finnish wind parks is created and tested.A neural network is trained and tested based on real turbine production data together with meteorological data. Customer provided data from two sites during 2013-2014 is used (Etha Wind operational data, 2014). The trained network can then be used in the estimation of icing for both different sites and years, since meteorological data is available for longer time periods and for many different locations in Finland.As the first step of the project, the network is trained to catch icing occasions in the data. Afterwards, the second step is to collect more production data, and decrease the uncertainty in the estimates. Finally, the icing occasion estimates are to be converted to icing-related production loss estimates.The network performed reasonably well in the test, 50 to 70 % of the icing was correctly classified. The number of observations classified as icing were very close to the actual amount of icing, 80 to 108 %. In conclusion, two significant improvements are found. First of all, the icing detection system of the parks was provento be unreliable, which causes problems in the neural network learning process. This system should be improved. Furthermore, more wind parks have to be tested to improve learning and decrease uncertainty in the estimates.
|Publication status||Published - 2014|
- wind power
- wind energy
- energy technology