Data Driven Early Stage Design Support for Offshore Wind Farms

This dissertation explores advanced methodologies in three critical areas of offshore wind energy development: data imputation for wind time series, predictive modeling of monopile-soil interactions, and machine learning applications in offshore ground investigation. The first study compares auto-regressive integrated moving average (ARIMA),, seasonal auto-regressive integrated moving average (S-ARIMA), and long short-term memory (LSTM) models for imputing missing wind data from offshore buoys, which are essential for early-stage wind energy assessments. LSTM models demonstrated superior pointwise accuracy, making them ideal for precise data reconstruction, while ARIMA was more effective in preserving the original data distribution. This comparison highlights the importance of choosing the right imputation technique based on the specific requirements of the analysis, whether accuracy or distribution preservation. The second study examines the application of artificial neural networks (ANNs) in predicting Cone Penetration Test (CPT) values across complex offshore geological conditions. CPT is an in-situ geotechnical method that uses a cone-shaped probe to measure soil properties like resistance and pore pressure as it is pushed into the ground. While ANNs proved effective in homogeneous deposits, their performance declined in heterogeneous environments, highlighting the limitations of current AI techniques in handling geological variability. The study suggests that AI-generated data is best suited for early-stage site characterization and survey planning, with more advanced methods required for detailed design. The third study investigates the use of artificial intelligence to predict monopile-soil interactions, focusing on the accuracy of AI-generated cone resistance values. The findings reveal that AI can closely mimic real-world behavior, making it a valuable tool for early-stage design, where rapid, cost-effective decision-making is crucial. However, the study cautions against relying solely on AI in later design stages due to potential deviations that could impact structural integrity, emphasizing the need for supplementary validation. By demonstrating the strengths of LSTM for data imputation, the conditional efficacy of ANNs in geotechnical predictions, and the utility—yet limitations—of AI in modeling monopile-soil interactions, this research advocates for a hybrid approach. AI proves indispensable for early-stage efficiency and cost reduction but must be complemented by traditional geotechnical methods in later design phases to ensure reliability, safety, and adherence to the complex realities of offshore environments. The dissertation also identifies key areas for future research, including improved imputation methods, refined AI models for monopile-soil interactions, and more effective machine learning strategies for complex geological environments.