Multi-model learning for vessel ETA prediction in inland waterways using multi-attribute data

Predicting vessel Estimated Time of Arrival (ETA) with accuracy and consistency is integral to Intelligent Transportation Systems (ITS), enabling reduced delays, enhanced operational efficiency, and more sustainable maritime logistics. Existing ETA prediction models largely rely on Automatic Identification System (AIS) data but often overlook additional factors. This study introduces a deep learning-based Multi-Model learning approach that fuses multi-attribute data from multiple sources to enhance ETA prediction accuracy. The model integrates Convolutional Neural Networks (CNNs) to extract spatial features, Long Short-Term Memory (LSTM) networks to model sequential dependencies, Transformer-based attention mechanisms to dynamically weigh environmental factors, and a Multi-Layer Perceptron (MLP) for incorporating vessel-specific and other residual features. The approach is evaluated on a large-scale dataset from the Weser River, an inland waterway with multiple locks, and benchmarked against traditional and deep learning methods, including K-Nearest Neighbors (KNN), XGBoost, MLP, and LSTM. Results show the Multi-Model achieves a Mean Absolute Error (MAE) of 39.37 minutes and a Root Mean Square Error (RMSE) of 69.58 minutes, representing an 81.50% improvement over the baseline. Analysis indicates that upstream movements and shorter segments yield lower errors. The findings highlight the importance of integrating spatial, temporal, and environmental factors for reliable ETA prediction.