The sensitivity to metocean data on using data-driven methods for a Valemax vessel speed prediction

Abstract

Accurate ship speed prediction is not only important for economic benefits and emission reduction, but also plays a vital role in making satisfactory routing plans. To investigate a speed prediction method of the largest bulk carriers ever constructed, we firstly compare several widely-used data-driven methods and then pick out the most appropriate model for the prediction task. Being as the input attributes of the data-driven models, the metocean data are collected from forecasting products and carefully cleaned up by data pre-processing steps. However, we have observed from literatures that the data for training are mostly collected from observational systems, such as the noon-reports, sensor measurements, on-board monitoring systems and reanalysis products, which may have certain differences with the data from forecasting systems. To figure out whether the data inconsistency problem will affect the model effectiveness on prediction, exhaustive experiments of a sensitivity study are carried out by utilizing observational data for training models while testing the prediction performance with forecasting data.