Vessel speed prediction using latent-invariant transforms in the presence of incomplete information

This paper presents a novel model designed to predict the vessel speed, specifically tailored to tackle the challenges posed by incomplete information of relevant operating parameters encountered in certain scenarios. In this method, a latent trend in the operating state of marine power system is firstly identified from historical time-series data to approximate the calm water speed information. Then, the modeling of the remaining component, which corresponds to the met-ocean-induced speed loss, can be more precisely targeted. Moreover, the elements situated at diverse temporal scales of the remaining component are disentangled, aiming to resolve the intricacies of coupled factor learning, thus improving the accuracy and validity of the model. For time-series with relatively steady-state, an LSTM network with a global attention mechanism is proposed to effectively capture the temporal evolution, and a differencing operation is incorporated to mitigate potential data inconsistencies between voyages. Finally, the proposed framework has demonstrated superior predictive capabilities for speed compared to a variety of data-driven methods, using a 400,000 DWT ore carrier as an example.