Abstract
Roll motion has a huge impact on operational efficiency and safety in the operations of marine vessels. Accurate roll motion prediction therefore serves as a critical component in decision support systems, enabling ship operators to plan and execute tasks efficiently while ensuring the safety of crew members, cargo, and equipment. However, prevailing roll prediction models fall short in terms of essential features for practical applications such as online learning and multistep forecasting. By recasting roll prediction into a long-term time series forecasting problem, recurrent neural networks within a sliding window framework are proposed, which can adapt in real-time and provide multistep predictions. This extended prediction window enables insights into roll motion oscillations, which are critical for accurate decision support. To evaluate the predicted roll oscillations, a new and robust metric is introduced to evaluate whether the model captures the true roll motion envelope. Real-world testing is performed to demonstrate the effectiveness of the model, highlighting its potential for improving decision support systems and improving operational safety.

This publication has 28 references indexed in Scilit: