Fatigue is often a governing design factor for offshore wind turbines. Since the design of offshore wind turbines includes conservatism, the actual accumulated fatigue damage can be lower than what the turbine is designed for. In this case, the operator can make a decision on life time extension of existing wind turbines. Therefore, it is important to estimate the actual accumulated fatigue damage to support decision making on life time extension, and for optimization of support structure design. However, fatigue critical locations are located near mudline where it is unfeasible to install strain gauges to measure the accumulated fatigue damage. The first purpose of this thesis is to investigate if data-driven approaches (linear regression and feed-forward neural network) can be applied to estimate the accumulated fatigue damage both in individual turbines and farm-wide levels. The second purpose is to determine the minimum number of sensors and quantity of data required for accurate estimation. Towards this goal, real measurement data of two offshore turbines in the same wind farm have been used. Specifically, the data-driven approaches have been applied with real measurement data from the SCADA system, measurements at the top and bottom of the tower, and data from a wave measurement system. This data was used to estimate the accumulated fatigue damage at multiple locations (tower bottom, transition piece and two levels on the monopile) in the form of damage equivalent loads. Throughout the study, 10 min. statistical properties of the measurement data have been used as input to the learning algorithms. One remark is that the estimation has not been performed for the fatigue critical location near mudline itself, but it is expected that estimation with these approaches can be expanded to the fatigue critical location if accurate response estimation at multiple locations on the support structure is possible. The results of this thesis show that the data-driven approaches can give accurate estimates damage equivalent loads on individual turbine level at multiple locations on the support structure when moment or inclination signals at tower bottom is used. For farm-wide level load estimation as well, it has been proven that the data-driven approaches can give quite accurate estimates the damage equivalent load. However, it should be noted that the turbines used in this study have similar dynamic properties. Therefore, the farm-wide level load estimation with the data-driven approaches should be further investigated in the future.

In the race for cost reduction in the offshore wind industry, support structure optimization leading to weight reduction plays a prominent role. The fatigue limit state is often the driving consideration for support structure design. Monitoring the monopile loads can offer an accurate knowledge of its consumed and remaining fatigue lifetime, which in turn has the potential benefits of lifetime extension and feedback on current design practices among others. Since within an offshore wind farm not all turbines are subjected to the same loading, a turbine or cluster-specific internal load monitoring scheme is investigated using data-driven approaches and specifically feedforward artificial neural networks (ANNs) and linear regression (LR).

The purpose of this thesis is to examine whether it is possible to accurately estimate the actual loading of the offshore wind turbine at the monopile mudline fatigue sensitive location by utilizing standard signals and/or sensor measurements and machine learning techniques instead of a structural model.

Data simulated with the Siemens Gamesa Renewable Energy in-house aeroelastic code - Bonus Horizontal axis wind turbine Code (BHawC) – is used, including operation and idling fatigue load cases. Ten minute time and frequency-domain statistics of signals that are collected in turbines are used as inputs to the data-driven models, whereas 10-minute damage equivalent loads (DELs) of monopile moments are used as targets. By using statistics of rotor rotational velocity, electrical power output, blade pitch angle, hub wind speed and nacelle accelerations, a mean absolute error of DEL estimation smaller than 4% in both operation and idling load cases is achieved, under the condition that the training set properly reflects the entire variation in environmental conditions. Furthermore, errors average out over multiple 10-minute intervals, resulting in accurate long-term estimation with residual errors between -1% and +1%. The accuracy of the estimation can be further improved by including additional sensors; accelerometers placed at the tower bottom (TB) can help reduce ANN mean absolute error by up to 40% in operational load cases.

In addition, if inclinometers are installed at TB they can allow 10-minute equivalent ranges of rotation signals at TB to be used as inputs to a LR scheme to accurately estimate DELs. When the 10-minute intervals are binned based on whether the mean wind speed is below, around, or above the turbine rated wind speed, this method gives mean absolute errors below 2.5%. As a result, TB inclinometers are recommended as extra sensors for load monitoring, provided that factors such as their accuracy level and durability compared to alternatives are considered.

The data-driven models examined in this thesis show potential not only for online internal load monitoring, but also for fast estimation of past load histories using recorded data.