Improving site-dependent power curve prediction accuracy using regression trees

The accurate prediction of the power production of a wind turbine at a particular site is important in both the planning and operation phases; however, the standard power curve binning method is not specific to the atmospheric conditions at the site. In this work, the application of machine learning for improving the accuracy of site-specific power predictions by taking into account turbulence intensity and shear is investigated, by creating a set of 8,000 ten-minute long aero-servo-elastic simulations of the NREL 5MW reference wind turbine at a random combination of hub-height wind speeds, turbulence intensities and shear factors using cloud computing with the software ASHES. A regression tree with maximum depth of eight and optimised with Adaptive Boosting is trained using a random selection of half of the data. For a set of 50 random test cases, the Root Mean Square Error of the predicted power compared to the simulated power is found to be three times smaller than for the standard power curve method of binning. OEMs could use this method to train a model for the dependency of power production on wind speed, turbulence intensity and shear in the certification phase of a given wind turbine type, which could then be applied by the wind farm planner or operator at any given site for which measurements of the atmospheric conditions are available. Similar success is observed for out-of-plane tip deflections. This result is highly dependent on the quality and distribution of the input data, and therefore this work is being used as a basis for the analysis of real measurement data, as well as for comparisons to other methods such as Artificial Neural Networks.