The research comprises the identification of the main drivers for wind turbine fatigue loads and tackles the question, whether these conditions are also reflected in standard wind turbine signals. Furthermore, possible techniques for the set-up of transfer functions to link standard signals and loads are investigated. Because of the many conditions which can potentially influence fatigue loads and the non-linear characteristics of wind turbines, a neural network based approach seems to be the most promising option, at least for loads encountered during normal power production.
Besides discussing the general outline of a load estimation system, different prediction schemes are developed from simulations of a typical 1.5 MW wind turbine. Loads and standard signals are generated using the industrial design tool Flex5. This allows to perform first test of methods for the prediction of equivalent load ranges and equivalent magnitudes from statistical parameters of standard signals. In addition, it is investigated if frequency domain methods can serve as the basis for the estimation of load cycle distributions. Because of the non-linear characteristics of wind turbines and the non-stationary operational conditions it turns out that these methods can not be applied easily. Therefore, an empirical method which employs fitting of combined probability density distributions to measured loads is developed.
Test and refinement of the methods is performed with measured data. In a first step the developed schemes are validated on the basis of data recorded from measurements at the state-of-the-art 5 MW wind turbine Multibrid M5000. The transferability of the approach to another turbine type and the feasibility to derive accurate predictions from a reduced number of input data is emphasised by additional tests at one 2.5 MW Nordex N80 turbine. First results indicating the applicability of already established transfer functions for load predictions at turbines in a series are derived by utilising measurements from a second N80 turbine. Finally, recommendations for further research activities towards the employment of a load estimation system on an industrial scale are given.