This master’s thesis project related inflow conditions to atypical fatigue loading of wind turbines. The research was fully based on experimental data obtained in another project. First an expected fatigue loading was defined based on the mean wind speed and the turbulence intensity. Differences between expected and measured fatigue loading were determined and analyzed using regression and clustering. The most important results regard the out-of-plane bending moment measured at the blade root and the side-to-side bending moment measured at the tower bottom. In addition to the mean wind speed and the turbulence intensity, the out-of-plane bending moment is found to correlate with wind shear. The side-to-side bending moment, on the other hand, correlates with wind veer and with different wind directions. In this case, the clustering result performs better in estimating the difference between expected and measured fatigue loading than the linear regression model.

The energy demand in current times has increased greatly in last few years. This increasing demand calls for a sustainable and clean energy resource that would reduce the load on non-renewable resources. Wind energy is a renewable resource which is harnessed by mankind from an ancient era. So as to meet this increasing energy demands, innovation in the field of wind energy is required.
A wind turbine generates electricity but achieving optimal power is a difficult task. In this case, the optimal power is defined as maximum power produced but at the constraints that the fatigue loading of the wind turbine structure should be as minimum as possible. Also, the wind turbine parameters such as rotor speed and pitch activity should be in a safe operational region. The problem in controlling wind turbines is that they work in a highly uncertain environment where managing so many factors at the same time are difficult. Optimal control of wind turbine has helped in achieving maximum power with-in safe working limits. Due to high uncertainty in the operating conditions of a wind turbine, it is quite a daunting task to find optimal gains for a wind turbine controller.
This thesis focuses on achieving optimal gain parameters for wind turbine controller by using an algorithm from machine learning community. In this thesis the problem is formulated as a supervised learning problem where input-output mapping has to be predicted. For this purpose, GPRT is used. The reason behind using GPRT is it takes fewer number of measurements to give good prediction compared to others. The property of GPRT where it deals with uncertain and non linear data with ease, making it a good choice for predicting wind turbine controller gains.
The second part of this thesis contains optimisation of the surrogate model achieved by performing regression. The optimisation is done by Monte Carlo Maximum distribution and improved results were generated by applying sequential sampling to this algorithm. This helps us to get a likelihood of optimal gains where the wind turbine gives out rated power with minimal fatigue loads, pitch activity and least deviation of rotor speed from rated.
The results obtained from the likelihood was tested for different operational wind speed and also tested for Extreme Operating Gusts as part of disturbance rejection and compared to current parameter used.
The comparison shows considerable improvement in the fatigue loads and pitch activity with having improvement in power production. In second case study, more parameters were predicted and optimised using the same algorithm so that the potential of this algorithm can be estimated. This was also performed successfully which proves that this technique can successfully be used to solve higher dimensional problems of wind turbine control.