Heerema Marine Contractors (HMC) is a contractor in the international offshore oil, gas and renewables industry. It is specialized in transporting, installing and removing large offshore facilities. HMC operates three crane vessels. Two of which are semi-submersibles (Thialf and Balder), the other is the monohull Aegir. A third semi-submersible, the Sleipnir, is currently under construction. To ensure safe operations, make accurate fatigue predictions and extend operational windows, HMC relies on vessel motion calculations. Currently, vessel motion estimations are based on response amplitude operators calculated by the diffraction software package WAMIT. As HMC cannot rely on diffraction software in case of non-linear vessel motions, the need for a method capable of capturing non-linear effects arises.

The goal of this study is to determine the potential of a neural network based model in predicting hydrodynamic behavior of semi-submersible crane vessels. Hindcast weather data, vessel motion measurements and model test data are used to train several different neural network architectures. The research into the potential of neural networks in predicting hydrodynamic behavior is split into two main categories: the frequency domain and the time domain.

Within the frequency domain, wave forecasts can be used to predict a response spectrum. The neural network in this case acts as a conventional RAO. In an artificial environment, four architectures are tested and the results show that neural networks are able to make accurate predictions in a fully linear environment. When tested on project data, where the vessel sails at operational draft, the neural network predictions shows a slightly higher accuracy than the diffraction based predictions for the specific test case. Another network is tested on transit data, where the vessel sails at an inconvenient draft. The results from these tests show that there is potential for a neural network to be used as a substitute for Response Amplitude Operators.

The time domain models focus on predicting ship response based on surface height signals and/or hindcast vessel motion measurements. The first model is trained and tested on model test data from an SSCV. The input of the neural network is surface height measurements and the output is pitch motion prediction. The model shows that it is capable of predicting both first and second order pitch motions. Another time domain model has MRU roll measurements as input and it tries to predict the future 60 seconds of roll motion. Many network topologies and optimizer settings are tested but none are capable of predicting future motions.