The potential gain in energy production and profit by improving the placements of turbines within a wind farm is driving interest. However, finding the optimal placements provides a complex problem. A large number of inter-dependent design variables create a design space that is difficult to solve. Therefore there is much attention for offshore wind farm layout optimisation (OWFLO) in practice and literature. Most of the research is done in selecting and creating the best optimisation algorithms, wake models and cost models.

This is not yet another study into better modelling or optimiser selection for OWFLO. Instead, this study aims to provide insight into what performance can be expected from OWFLO and to know when further optimising is not justifiable anymore. The study consists of three parts. All three parts make use of a referent. A referent can be considered a close representation of reality, obtained by a best-practice implementation of the optimisation problem and its associated models. It is assumed that the referent has the same characteristics as reality and that deviations of other implementations of OWFLO from the referent are representative of their deviations from the true optimisation problem.\\

In this study, the referent is defined by, amongst others, the use of the Bastankhah and Porté-Agel Gaussian wake model, a 12 sector wind rose with a Weibull distribution per sector and the gradient-free covariance matrix adaptation evolution strategy. The referent is used to maximise annual energy production (AEP).

The first part uses the referent to find and understand the characteristics of the OWFLO problem. Wind farms with 9, 25 and 64 turbines have been optimised 100 times with the referent. The results show a small spread in the performance of the found optimised layouts, indicating that many local optima exist with similar performances in an OWFLO problem. The spread between the highest and lowest found performance decreases with increasing numbers of turbines. A special form of the response surface is used to visualise the response surface. The visualised response surfaces, with only two design variables, showed that the wakes of the turbines created multiple local optima.

The second part compares performances from optimised layouts with 25 turbines resulting from optimisations with alternative implementation choices, evaluated by the referent model. The performances are represented in boxplots for 100 optima each. The boxplots show that the influence of alternative implementation choices depends on the slightly different locations of the local optima in the design space and the roughness of the response surface they create. The influence of the shifts in locations of the local optima on the performance turned out to be minimal. An increase in the roughness of the response surface meant an increase in the spread of the performances. The difference in performance resulting from the alternative optimisers indicates that improvement of a state-of-the-art optimiser is not expected to lead to much better results.

The third part explores the need for improvement of the analysis by adding a phenomena currently not considered in OWFLO. The influence of neighbouring wind farms on layout optimisation without including atmospheric stability is explored. Three cases have been defined to show the influence of neighbouring wind farms on layout optimisation. It is concluded that adding neighbouring wind farms for accurate energy yield assessments is necessary. However, for layout optimisation, the benefit of including neighbouring wind farms is not evident.