A novel maritime traffic complexity model based on ship trajectory prediction

As global shipping density increases, maritime traffic in complex waterways exhibits pronounced dynamic and multi-scale characteristics, while traditional static assessments struggle to support proactive risk warning. To address this gap, this study proposes a novel, unified framework that integrates prediction, clustering, micro-level complexity quantification, and Shapley-based cross-scale fusion into a single methodology. First, a Bidirectional Gated Recurrent Unit Sequence to Sequence (Bi-GRU Seq2Seq) model with a fusion attention mechanism is constructed to achieve short term vessel trajectory prediction based on AIS data. Second, predicted trajectories are clustered using the density based spatial clustering of applications with noise algorithm to identify potential encounter clusters. Third, a micro level complexity model is developed from four dimensions, namely motion trend, relative distance, distance at closest point of approach, and relative bearing, to quantify vessel interaction risk. Finally, Shapley value theory is utilized to realise nonlinear cross scale fusion from vessels to clusters and to the regional level, enabling marginal contribution analysis and interpretable macro level assessment. A case study in the Laotieshan Channel of the Bohai Sea demonstrates that the prediction model achieves a Root Mean Square Error of 84.99 m, with a maximum error reduction of 76 per cent compared with baseline models. The proposed framework effectively identifies risk hotspots and captures the aggregation and diffusion patterns of traffic complexity. By enabling a transition from static assessment to proactive risk warning, this study provides a quantifiable and interpretable methodological basis for intelligent maritime supervision, refined vessel traffic services management, and proactive traffic risk warning. • Establishes a prediction-driven multi-scale framework shifting maritime traffic assessment from static description to dynamic evolution analysis. • Integrates AIS-based trajectory forecasting, encounter clustering, and complexity modelling into a unified computational chain. • By integrating a fusion attention mechanism into the Bi-GRU Seq2Seq architecture, the proposed model achieves an 84.99 m RMSE. • Proposes a four-dimensional micro-level interaction model and introduces Shapley-based nonlinear cross-scale complexity fusion with interpretability. • Validates the framework in the Laotieshan Channel, revealing risk hotspots and aggregation–diffusion dynamics of traffic complexity.