Implementation of machine learning to model losses from icing of wind turbines

Abstract: This thesis investigates the possibility to use machine learning algorithms to predict the losses due to icing in the Stor-Rotliten wind farm that is situated in the north of Sweden and operated by Vattenfall. The inputs for the machine learning are historical mesoscale modelled variables that are derived from a Weather Research and Forecasting Model code that is tuned for icing (WRF-model). An ice model has been updated and improved so that it would achieve a better indication of icing, based on the equations from Lasse Makkonen. A more accurate model of a wind turbine considers the turbine blade as a rotating cylinder at 85% of the length of the blade and not as vertical cylinder that stands still. Besides this, the variables from the mesoscale data are used as inputs for the machine learning algorithm. The targets are energy production losses due to icing that is computed from historical SCADA data that covers the same time frame as the WRF data. To reduce the complexity and the computational time of the system a statistical variable selection algorithm, called mutual information, is used with the MILCA toolbox for Matlab. The target for the variable selection and the machine learning is the average loss of power per wind turbine per hour. This is extracted from the production data from Vattenfall. The goal with the thesis is to relate the modelled mesoscale data with the production data (SCADA). The overall result of the study is that the neural network method offers a suitable and more accurate way to predict the losses from icing on wind turbines, but there is some work still to be done to reduce the errors in the input variables.