MOORING OPTIMIZATION USING ML TECHNIQUES

The mooring optimization process for a FOWT is a complex study that involves a large number of simulations. All possible designs of the optimization process should be assessed and the operational and survival constrains verified for acceptance or disregard the solution. These constraints are mainly the design bases of the mooring lines, but also the FOWT operational and survival conditions, which are mooring system dependant as it is the governing component for surge, sway and yaw stiffness. Machine Learning (ML) algorithms are used to predict the simulations results by training the model from a set of defined simulations, allowing for a significant reduction of the computational cost of the large number of required simulations. The methodology's main advantage is the high velocity once the model is trained. However, some uncertainties can arise from the exactness of the predicted values. The aim of the paper is to design an optimized mooring system for the WindCrete platform for a Gran Canary Island with a 200 m sea depth for ULS. To overcome the difficulties of the optimization process two ML models are developed. The first one, the static ML model, is based on the static response of the mooring system that allows to reduce the solution space of the optimization problem. The second one, the dynamic ML model, is based on the dynamic response of the FOWT and allows to assess the constraints applied to the objective function as penalties. The objective function is defined as the total mass of the mooring system. The variables of the optimization problem are the main line and delta line lengths, the radius to anchor and the chain diameters. The first ML model is set-up to predict the static mean line tension at rated wind speed, the vertical force on anchor at rated wind speed and the initial Yaw period of the FOWT. These parameters allow to discard a large number of possible solutions that do not fit with the following design bases: the maximum line tension and no vertical force on anchor. Moreover, a maximum yaw natural period threshold is set-up based on design experience. This is needed due to the lack of damping in yaw direction of Spar platforms. The static ML model is used to create an initial sample of feasible solutions to train the dynamic ML model. Also, during the optimization process is used as a classification model to exclude non-feasible solutions to ensure the performance of the second ML model. The second ML model is based on the dynamic response of the WindCrete platform and the mooring system using OpenFast simulation tool. The model is set-up to predict the maximum line tension, vertical force on anchor, maximum surge position and maximum pitch which are used as constraints parameters to be applied at the optimization function as penalties. The results show a good approximation of both ML models with a high potential to be applied in determining design load cases, including fatigue assessment in the optimization design process.