Author(s): Robert Burton; Renat Yulmetov; Rocky Taylor
Linked Author(s):
Keywords: Lake Ice; Ice Mechanics and Properties
Abstract: Offshore structures in ice-prone areas are often designed with sloping elements at the waterline to promote ice failure in bending, in order to reduce the total ice load. In this case, the load calculations are parametrized using flexural strength of ice, which is scale and temperature dependent. Flexural strength of ice has been measured in a variety of experiments, both in laboratory and in situ. A comprehensive database of beam flexural strength measurements for fresh water ice was acquired, consisting of more than 2000 records, and included corresponding details on beam volumes, ice temperatures, and locations. In this work, machine learning (ML) regression algorithms are developed and applied to predict flexural strength of fresh water ice. Using several of the most common ML algorithms, including multilayer perceptron, gradient boosted trees and Gaussian processes, flexural strength models are parametrized. These models are compared and contrasted looking at model bias, accuracy and generalization as well as their ability to extrapolate and predict strengths for values outside of the training data. The application of ensemble models is also investigated, in which two or more models are blended, attempting to overcome their individual weaknesses by building upon their combined strengths. These models are applied in probabilistic design load calculations and are compared with the conventional approach for a wind turbine installation in Lake Erie. In the conventional approach, the flexural strength is generated probabilistically, while in the ML model the strength is a function of temperature and volume. The implications of implementing the ML approach on uncertainty in 50-year design level ice loads on a downward breaking cone of the wind turbine are discussed.
Year: 2022