Abstract

The maritime industry is responsible for the vast majority of the worlds goods, making modest energy efficiency improvements have a big impact in minimizing emissions and reducing operational costs. Accurate forecasting of propulsion power in marine vessels, as a tool for exploration of optimal voyage configurations, is crucial for optimizing fuel efficiency. This study explores time-series forecasting techniques to predict propulsion power consumption based on weather data and on-site measurements from two roll-on/roll-off passenger (ROPAX) vessels. We analyze the raw data to understand the relationships between features through a statistical analysis, establish a theoretical framework that allows for domain knowledge integration and reinforcement learning compatibility to the forecasting process, and compare different modeling options regarding their accuracy and scope. The findings highlight the potential of the proposed forecasting method, showing that it is feasible to forecast propulsion power for an upcoming voyage. Out of the analyzed models, Long Short-Term Memory (LSTM) networks and Extreme Gradient Boosting (XGB) work best to predict the behavior of the vessel's environment, with XGB being the most robust to sparse input data, higher output dimensionality and outliers, while LSTM networks provide higher accuracy with more strictly controlled input data.