Application of Ensemble Methods for Solving Offshore Wind Farm Layout Optimization Problems
MetadataShow full item record
Wind energy is considered as one of the most promising renewable energy sources. The intent of this dissertation is to assess the application of ensemble methods on offshore wind farm layout optimization. In this regard, four research questions are considered and addressed accordingly each of which are discussed in detail in a separate chapter (chapter 2 to 5). The first research question is focused on choosing appropriate analytical wake model to be used for the wind farm layout optimization (WFLO) problems in this dissertation. The WFLO problem is a complex and non-convex optimization problem. The accuracy of analytical wake models applied in WFLO problems is of great significance as the high-fidelity methods are still not able to handle an optimization problem for large wind farms. Four wake models selected from FLORIS (the abbreviation of Flow Redirection and Induction in Steady State) which is a tool integrated with several classical and innovative wake models by applying three classical WFLO scenarios are compared. The results illustrate that the Jensen wake model is the fastest, but the issue of underestimating the velocity deficit is obvious. The multi-zone model needs additional tuning on the parameters inside the model to fit specific wind turbines. The Gaussian-curl hybrid (GCH) wake model, as an advanced expansion of the Gaussian wake model, does not provide a significant improvement in the current study, where the yaw control is not included. The Gaussian wake model is recommended for the WFLO projects implemented under the FLROIS framework and has similar wind conditions with the present work. As the analytical wake model has been confirmed, the second research question is focused on a comparative study for multi-stage optimization models which are proposed to enhance the performance on searching the optima and convergence speed. Even though many different heuristic algorithms and mathematical programming methods have been tested and discussed, there is not a consensus about which algorithm is the most suitable approach for solving WFLO problems. Every algorithm has its own advantages and disadvantages on solving different problems, thus the ensemble approaches have received attentions. One ensemble approach applied in solving WFLO problems is to apply the multi-stage model as an algorithm in stage 1 to capture a coarse, initial optimized layout and import it to stage 2 as an initial condition for another algorithm for further refinement. Two types of multi-stage methods are compared: The Heuristic-Gradient-based (H-G) model which consists of a heuristic algorithm in stage 1 and a gradient-based algorithm in stage 2; The Discrete-Continuous (D-C) model which consists of a heuristic algorithm in discrete scheme in stage 1 and an algorithm in continuous scheme in stage 2. Annual energy production (AEP) is used as the objective function while the computational time associated with each approach is documented. The results illustrate most of the multi-stage models can improve the optimization procedure both in terms of AEP and computational time. Overall, it is found that the D-C approach is better than the H-G approach. Particularly, the combination of Greedy and Random Search provides the highest AEP and the combination of Greedy and SLSQP provides the lowest computational time. Despite of which approach is applied, solo algorithms or multi-stage models, the optimization relies on wake calculation which consumes the most of computational time. Thus, the third research question is about applying machine learning technique to construct a surrogate model for wake calculations in solving WFLO problems. An approach by neural network (NN) is developed and discussed. The NN can predict the site AEP by inputting the wind rose data and site turbine density. The dataset consists of randomly generated samples as 70% of them are used for training and the remaining samples are used for testing. The dependencies of the prediction quality in order of the variations in hyperparameters are also addressed by comparing different network architectures. The trained NN presents 96.78% as the highest mean accuracy in the testing task. This approach can be applied by an entity when making decisions of selecting potential locations for constructing offshore wind farms as an initial screening process. It can also be applied as the algorithm in the first stage of multi-stage model for solving the WFLO problems containing large number of wind turbines where applying the classical optimization algorithms are still inoperable due to the time consuming. Building floating offshore wind farm in the deep waters is a future demand of renewable energy. Because of the six degrees-of-freedom involved in the concepts, the load analysis should not be omitted in the WFLO for floating offshore wind farms. Finally, the fourth research question is about the joint optimization analysis for floating offshore wind farms with considering transient ambient wind and turbine structural loadings. A small wind farm (wind turbine array) which consists of two 15MW floating offshore wind turbines (FOWTs) and two 5MW FOWTs is investigated. The semi-submersible floater is adopted. The wake meandering, power production, rotor torque and selected platform motions (surge and pitch) are observed based on the instantaneous simulations in three wind speed scenarios. The simulations are completed by NREL’s FAST.Farm software through parallel computing. The observations reveal that wake meandering has strong impacts on the power performance and dynamic responses when the wind speed is low and the distance between turbines is small. The configuration of ABBA type, i.e., two 15MW FOWTs on the sides and two 5MW FOWTs in the middle is more stable than the other as ABAB in power generation and structural responses. In the study, the resonances on platform surge and pitch are observed in the low wind speed scenarios, but they are not functional to the distance between turbines.