Evaluating statistical learning approaches to predict suction bucket displacement due to vertical cyclic loading in sand

Bottom-fixed offshore wind turbines (OWTs) are sensitive to tilting, and therefore strict out of verticality limits apply to the foundation serviceability limit state design. In the case of OWTs supported by suction bucket jackets (SBJs), tilting of the turbine arises from differential displacement between the windward and leeward suction buckets, such that predicting the foundation displacement due to vertical cyclic loading is critical. Assessing displacement accumulation due to cyclic loading for all the suction buckets in an offshore wind farm would require significant time, and detailed soil and loading information may not be available at the early stages of the design. Statistical learning models can map complex non-linear interactions between features and target variables by performing regression techniques in a given dataset (training data). Once the patterns in the training dataset have been learned, predictions can be performed orders of magnitude quicker than numerical models, making them well suited to design practice, particularly during the feasibility stage of a design. This paper investigates the performance of three non-linear statistical learning models (General Additive Model, Random Forest and eXtreme Gradient Boosting) in predicting the accumulated displacement of suction buckets due to vertical cyclic loading. The research data are taken from centrifuge model tests that feature over 80,000 load cycles with varying mean stress, stress amplitude and drainage conditions. The model performance was assessed using statistical metrics (coefficient of determination and mean squared error) and by comparing the measured and calculated displacements for storm loading, with the ensemble models providing encouraging results. The prediction making process of the best performing model (eXtreme Gradient Boosting) was investigated using a game theory approach (SHappley Additive exPlanations) and was shown to be consistent with current engineering knowledge. Notably, the best performing model was able to capture the effects of changing stress amplitude during storm loading, offering a more realistic representation than the cyclic load magnitude ordering approach that is typically adopted in engineering practice.