Ship motion forecasting under varying operating conditions via multi-feature fusion

Ships navigating in the ocean are subject to irregular perturbation motions. Accurate forecasting of such motions significantly enhances navigation safety and onboard operational efficiency. However, traditional ship motion forecasting methods are typically designed for single, stable operating conditions and lack generalizability across varying conditions. To address this limitation, we propose Multi-Feature-Informer (MF-Informer), a motion forecasting model based on multi-feature fusion technology, applicable to most operating conditions. The model is trained and evaluated on a dataset comprising perturbation motion data of the KVLCC2 vessel under 1000 randomly distributed operating conditions in sea states 2–6. It employs band-pass spectral extraction techniques based on prior physical knowledge of ship spectral distributions when extracting frequency features. Multi-feature fusion techniques, including Cross Attention and concatenated linear projection, are employed and compared in this study. The model’s hyperparameters are optimized using the Sparrow Search Algorithm (SSA). Experimental results demonstrate that, compared to models without feature fusion, MF-Informer reduces the mean squared error (MSE) for heave, roll, and pitch forecasts by $24.29\%,20.24\%,26.27\%$, respectively. Additionally, the operating condition feature extraction module exhibits strong physical interpretability.