Prediction of vessel propulsion power using machine learning on AIS data, ship performance measurements and weather data

Abstract

The aim of this paper is to utilize AIS (Automatic Identification System) data, ship propulsion power measurements and weather data and apply different machine learning (traditional and deep learning) methods to develop improved models to predict ship propulsion power. The performance between different traditional machine learning methods and deep learning with different architectures were compared and discussed. Two scenarios were explored: 1) Training a machine learning model on one container ship significantly improved the predictions of power demand as compared to a physics-based model, with an R2 score of 0.78 compared to 0.48. 2) A machine learning model was also trained on several container ships. This scenario, where the test data with other container ships was not included in the training dataset, also showed better prediction with machine learning than physics-based model, with an improvement in R2 score from 0.69 to 0.85. However, the use of the trained machine learning model on other ship types showed varying results and especially when the vessel size differed significantly from vessels in the training data, data-driven models showed limitations. For vessels of similar size, however, machine learning for other ship types showed improvements.