Convolutional neural network with dual inputs for time series ice prediction on rotor blades of wind turbines
Markus Kreutz, Abderrahim Ait Alla, Kamaloddin Varasteh, Jan-Hendrik Ohlendorf, Michael Lütjen, Michael Freitag, Klaus‐Dieter Thoben
Abstract
Downtimes due to ice formation on rotor blades reduce the economic efficiency of wind turbines. An accurate ice prediction is required to operate active de-icing measures such as blade heating as an anti-icing system. Building upon our previous research, this paper proposes the use of a convolutional neural network model with dual inputs and one-dimensional convolution filters using historical data from the wind turbine as well as weather forecasts to predict the ice situation for the next 24 hours. The model is validated using data from three different wind farms and shows an average balanced accuracy of 97.9%.