Deconstructing spatiotemporal complexity in Inland shipping: Vessel traffic flow prediction based on hexagonal grids and adaptive mixture of experts architecture

Accurate forecasting of inland vessel traffic is pivotal for waterway safety. It remains a challenging task due to the significant non-linearity and spatiotemporal variability inherent in maritime systems. Existing approaches reveal shortcomings, as they struggle to decouple complex dynamics. In particular, conventional spatiotemporal graph neural networks (STGNNs) rely on static structures that fail to differentiate between fixed physical constraints, which are primarily governed by topology and meteorology, and transient dependencies induced by congestion. To address these limitations, we propose the Mixture-of-Experts Vessel Traffic Flow Prediction (MoE-VTFP) framework, which decomposes the prediction landscape into distinct dynamical sub-problems. By employing an H3 hexagonal grid system, we minimize spatial anisotropy. Our architecture employs a gating network to adaptively synthesize expert modules specialized in long-range temporal dependencies, static environmental topologies, and dynamic flow interactions. Empirical validation using Xunjiang River Automatic Identification System (AIS) data demonstrates that MoE-VTFP exhibits robustness across diverse hydrological conditions. Our experiments results indicate that the model maintains a Mean Absolute Percentage Error (MAPE) consistently below 9%, with a minimum of 5.65%, and a coefficient of determination ( R 2 ) exceeding 0.93 during our experiments. By effectively handling seasonal volatility, this framework offers a rigorous foundation for next-generation intelligent waterway management.