Improving knowledge and forecasting of ship performance in waves via hybrid machine learning methods

Two data-driven hybrid machine learning architectures are presented to improve knowledge and forecasting capabilities for ships operating in waves. These are based on methodological extensions of both dynamic mode decomposition (DMD) and recurrent-type neural network (RNN). Namely, a full-rank DMD approach is augmented by the use of time derivatives and timeshifted copies of time histories, while RNN is applied to sequence-to-sequence learning and uses long short term memory and gated recurrent unit layers. The two architectures proposed here combine DMD and RNN using a parallel and a serial approach, respectively. Results are presented and discussed for the course keeping of the 5415M model in stern-quartering sea state 7 irregular waves at nominal Fr = 0.33. DMD provides ground for physical interpretation on the ship dynamics via modal representation of the interconnected variables (such asmotions, forces, moments, etc.) based on the observed time series. Furthermore, it is also able to accurately forecast ship performance for short temporal horizons (about one encounter-wave period). RNN provides no direct physical interpretation, but is found more robust to forecast ship performance for longer temporal horizons (three encounter-wave periods). The most promising hyperparamteres sets for DMD and RNN are identified for the current test case, based on a full-factorial combination of setting parameters tested against a random sample of sequences, using four evaluation metrics. The hybridizationof DMD and RNN provides a viable approach to accurate and interpretable predictions, as shown and discussed in the paper.