On equivalent ice thickness and machine learning in ship ice transit simulations
MetadataVis full innførsel
- Institutt for marin teknikk 
This thesis presents a contribution to the research field of ship ice transit simulations used for estimating the ship resistance/speed profile in a complex ice field. Namely, two distinct methodological approaches within the field are addressed: equivalent ice thickness-based methods, and machine learning-based methods. Contrary to the computationally expensive high-fidelity mechanistic numerical models, both of these have an intrinsic characteristic of low computational costs, making them potentially attractive in the upcoming age of autonomous shipping, when the operational decisions will have to be made constantly and automatically in very short time intervals based on a continuous data stream of dynamically changing environmental parameters. The concept of equivalent ice thickness utilized for the simplification of a complex ice cover, although widely used within studies of shipping in ice, has not been thus far thoroughly studied and validated. Therefore, this thesis presents a first complete and systematic overview and validation of the existing definitions of equivalent ice thickness. Moreover, a novel method for the calculation of equivalent-volume ice thickness is developed, not requiring ridging parameters as an explicit input, showing an increased accuracy in the estimation of the total volume of ice along a sailing route by reducing the error from 29% to 2%. Finally, the traditional concept of equivalent ice thickness based on equivalent ice volume is extended, and the concept of equivalent-performance ice thickness is developed, resulting in the reduction of error of simulated ship speed of up to 70% compared to the traditional approach. Recently, data-driven approaches based on machine learning have gained increased attention within different fields of engineering, providing a computationally efficient and potentially more accurate alternative to the traditional mechanistic simulations. Therefore, this thesis presents a prototype of the first simulator of ship speed profile in a complex ice field based on machine learning. The results show a promising accuracy of the developed method, with an average error of simulated ship speed against the measured one ranging from 2.6% to 9.4%. The thesis concludes that the developed methods present valuable tools for the simulation of ship performance in ice, considering both their accuracy and the low computational costs. Moreover, they can be used in conjunction where the equivalent ice thickness-based method is suitable for voyage planning on a strategic level, while the expected ship speed profile on a tactical level can be predicted using the developed machine learning-based method.