Predicting estimated time of arrival for ships: A frequency-based approach considering met-ocean factors

• A framework is improved to more accurately predict ship estimated time of arrival. • Key information is extracted from original speed signal by a well-designed strategy. • Sub-models are built based on distinct characteristics of corresponding components. • Experiment results from practical case study verify the superiority of our framework. Accurate prediction of the estimated time of arrival (ETA) of ocean-going ships facilitates coordinated dispatch between ships and ports, alleviating congestion and enhancing efficiency. For ships navigating along predetermined routes to complete scheduled voyages, ETA prediction methods are broadly classified into direct and indirect models. Compared to the former, which directly infers ETA from potential influencing factors, the more practical indirect methods divide the problem into ship speed prediction and geographical distance calculation. However, existing models, which attempt to learn from the original speed signal with mixed information, struggle to capture ship navigation characteristics under complex operational and hydro-meteorological conditions. Hence, we propose an indirect framework for more accurate and robust ETA prediction, in which the speed signal is decomposed into multiple sub-signals with distinct characteristics. Specifically, for the nearly invariant sub-signal, a lightweight model is used, while a novel deduction process is designed to estimate the nearly periodic sub-signal, leveraging its inherent stable frequency information. In addition, a multi-feature semi-dynamic model is employed to simulate the high-frequency sub-signals, owing to its sensitivity to the external environment. Finally, our proposed framework demonstrates superior performance over other state-of-the-art methods in practical case studies, achieving a minimum 30 % reduction in prediction error.