The greenhouse emissions caused by the energy sector must be significantly reduced to slow down the impact on climate. Therefore, a great shift towards renewable energy sources is of importance. A method of doing so is by using offshore wind energy. The development and growth of this renewable energy source introduce new challenges for the future. One of these challenges is designing offshore wind turbines for sea-ice loads. In Europe, the northern Baltic Sea is the main site of interest for the development of offshore wind farms (OWFs) that will be subjected to ice loads. However, little is known regarding how the ice behaves around these structures and how the presence of multiple turbines will influence the ice loads encountered. Therefore, the following research question is defined.

How is the ice loading on and ice drift around an offshore wind turbine influenced by it being
part of an offshore wind farm?

First, an introduction to sea-ice terminology and ice action is given. This forms the theoretical framework for this thesis. Multiple methods are explored to analyse and summarise the ice loads experienced by a structure. These are later used to interpret the results obtained from simulations. In this thesis, the main area of interest is the northern Baltic Sea (Bay of Bothnia). Therefore, the environmental conditions and sea-ice in this area have been thoroughly investigated. Image processing of ice fields over the years yields a starting point for generating ice fields that are used in the simulations.

Furthermore, all relevant ice properties are considered and, when possible, adjusted for the lower salinity present in the Baltic Sea water. In addition, current and wind conditions are determined to determine in which way these play a role in the drift of the ice.

In this thesis, both the effect of the wind directions and ice concentration are investigated. Combining these results in 16 total simulation scenarios that will be modelled using the Simulator for Arctic Marine Structures (SAMS) software. In total, four different wind directions (0, 15, 30 and 45 degrees) and four different ice concentrations (50%, 60%, 70% and 80%) are used. The OWF used has a grid layout with 5x5 turbines and an ice field of 15x15 km. The ice field is simulated such that it drifts into the OWF, no ice is initially present in the wind farm. Due to computational limitations, for each scenario, 45000 seconds (12.5 hours) of data is simulated.

To analyse the obtained force data, the Ice Load Factor (ILF) and the Ice Load Contact Factor (ILCF) are defined. Both use rainflow counting to analyse the number of force cycles that occur over time. A lower threshold of 100 kN is used to focus on the higher impact interaction. In addition, an upper limit of 10.8 MN is used to eliminate numerical peaks from the results. Both factors use the sum from the product of each force bin and it’s corresponding amount of cycles. Each load bin has a size of 0.5 MN and the average value of this bin is used in the product. Summing this over all load bins yields the numerator for both factors. The main difference between the factors is the denominator. For the ILF the total potential exposed time is used based on the free-drift velocity of the ice. Since this does not take into account the actual time that the interaction occurs, the ILCF uses the total contact time of the ice with the structure. Both fractions are multiplied by the total simulated time to obtain a unit of Newton in both factors.

Based on the obtained data one can analyse each individual turbines and also group them. In this thesis the term, ”Lines of turbines” is used to refer the order in which the ice interacts with the OWTs. Hence, the first line of turbines represents the turbines that first interact with the ice field. Moving further downstream, the following up lines are defined and can be analysed for all wind directions. Analyses on these lines of turbines gains insight on how the ice action changes when moving further down the OWF.

For the lowest ice concentration (IC = 50%) the ILF decrease for all wind directions in the lines of turbines is approximately the same. Only when the ice concentration increases, a more clear difference is seen for changing wind directions. Since a higher IC results in more ice floes in the ice field. Therefore, it is more likely that a turbine interacts with an ice floe. This influences the drift in and around the wind farm. Generally, more sideward movement is caused by higher ice concentrations. This makes ice more likely to interact with turbines further downstream. For lower concentrations, the ice continues to move along the same streamline resulting in less ice-structure interactions. The mean ILF is decreased by >70% and the sum by >75% after interacting with the first line of turbines in all scenarios. Hence, a significant reduction is present in the wind farm.

The interaction times for a 0 and 15 degree heading decrease significantly more over the lines of turbines. Furthermore, these wind angles result in the lowest ILCF sum and mean values for all ice concentrations. Based on the ILCF, one can conclude that these two wind directions result in the best reduction of ice loads based on the total interaction time. Furthermore, one concludes that the 30 degree wind angle yields the lowest reduction in interaction time and therefore has the lowest ice-structure interaction blockage effect. This could potentially be caused by the fact that this angle has the highest offset in the position of the turbine in y-direction. Therefore, turbines in previous lines do not prevent other turbines further downstream to interact with the ice field.

The computational times of these kinds of simulations are found to be significant. Therefore, it is explored whether machine learning (ML) could be used to predict the encountered ice loads. This could greatly influence the pace with which different OWF layouts or ice conditions can be tested more thoroughly. From the data obtained, different ML regressors are trained and tested for their performance. The usage of a Random Forest Regressor resulted in the most accurate results. However, these result are found not to be accurate enough to use them for prediction purposes. Therefore, it is found that first more data need to be gathered to train and test these models more extensively, and the usage of more complex ML models could potentially also improve the performance. Doing so could greatly influence the speed with which different parameters of influence can be analysed.