Development of a dynamic deep learning framework for the prediction of vessel arrival time to port

Dynamic vessel estimated time of arrival (ETA) to port is crucial for port operations and the efficiency of the global logistics system. However, existing studies on vessel ETA prediction often lack detailed descriptions and analysis of the datasets used and fail to provide dynamic predictions at different stages of a vessel’s voyage as the real-time ship trajectory data accumulates. To address these issues, this study proposes an innovative hybrid voyage-adaptive ETA (HVA-ETA) prediction framework to facilitate the utilization of both vessel static features and voyage dynamic features for vessel ETA to port prediction. The static features are learned by a tree-based model and the dynamic features are learned by a baseline model based on deep learning (DL), which iteratively predicts the vessel’s trajectory to the next port of call. The predicted trajectories are processed as intermediate features and further combined with observed dynamic features to serve as the inputs of the subsequent ETA prediction model. To comprehensively assess ETA prediction performance across different voyage stages, we propose a set of novel evaluation metrics, covering aspects of prediction accuracy, monotonicity, and stability. We conduct experiments using AIS data from container ships traveling from four ports in the Guangdong–Hong Kong–Macao Greater Bay Area to the Port of Singapore in 2021 to validate the performance of the proposed framework. The results demonstrate that the proposed HVA-ETA framework consistently outperforms baseline deep learning models (RNN, LSTM, GRU, Autoencoder, and Transformer) and five existing ETA prediction approaches. In particular, across different voyage progression stages and shipping routes, HVA-ETA achieves an average reduction of approximately 1.33 h in prediction error compared with current practice.