Analysis and prediction of research vessel attitude based on full-scale operational data

Abstract

This study uses full-scale operational data from a large research vessel to perform statistical characterization and guide the selection of data-driven prediction models. A simplified Long Short-Term Memory-based (LSTM-based) framework, referred to as LSTM-FC, is developed for computationally efficient multi-step attitude prediction of research vessels. Initially, the measured pitch and roll angle data recorded during a transoceanic voyage, collectively referred to in this study as vessel attitude data, are analyzed using the fast Fourier transform and the autocorrelation and partial autocorrelation functions, revealing pronounced temporal dependence and non-stationarity. Building upon these insights, prediction frameworks based on LSTM and sequence-to-sequence (Seq2Seq) LSTM are developed. Motivated by the Seq2Seq LSTM structure, the LSTM-FC framework replaces the recurrent decoder with a fully connected layer, thereby reducing computational iterations and training latency while maintaining competitive predictive performance. For a 10-second prediction horizon, the LSTM-FC model achieves an RMSE, MAE, and Pearson correlation coefficient r of 0.1385 ± 0.0003, 0.1021 ± 0.0004, and 0.7694 ± 0.0011 for pitch angle prediction, and 0.1521 ± 0.0037, 0.1119 ± 0.0039, and 0.7357 ± 0.0166 for roll angle prediction. The maximum effective prediction horizon of the LSTM-FC framework is further evaluated.