Robin Marcille thesis defense

Offshore wind energy maintenance operations are highly sensitive to forecast uncertainty. Numerical weather prediction are limited by their computational cost for the uncertainty estimation and the update frequency, which is an argument for the development of data-driven methods. The importance of offshore measurements is highlighted by the results. A method for designing an optimal sensors network is proposed using unsupervised clustering. This method has been used by the French weather service to define future networks of floating LIDAR for offshore wind. Deep learning models for the joint probabilistic forecasting of metocean parameters are proposed. Their relevance for assimilating a large amount of input data is demonstrated. A Gaussian posterior and a generative approach using normalizing flows are compared. It is shown that the use of normalizing flows can relax any assumption on the shape of the forecast probability density while maintaining sampling and likelihood computation capabilities. A real case study dataset is built on a relevant area for offshore wind. The probabilistic models are adapted for joint wind and wave forecasting, for which the non-Gaussian properties of the normalizing flows is beneficial for forecast reliability. An evaluation framework dedicated to offshore operations is proposed, including the generation of probabilistic scenarios and the measure of decision-making economic impact. It is shown that the search for an economic optimum in the probabilistic decision-making leads to higher risk during operations, and this should be taken into account for forecast selection and evaluation.