Robust Feature Fusion Perceiver IO-GNN for Digital Twin Quay Crane Scheduling in Smart Ports

Source abstract

Quay-crane scheduling in smart ports must accommodate irregular vessel arrivals, unreliable sensor streams, yard-side congestion, and safety separation constraints that are difficult to represent with static dispatch rules. This study develops a robust feature fusion Perceiver IO-GNN model for digital-twin-based quay-crane scheduling. Vessel, crane programmable logic controller, berth, yard-truck, and weather observations are first aligned in a reliability-weighted feature space that retains missingness, observation age, and cross-source consistency. Perceiver IO then maps the asynchronous multimodal tokens into a fixed latent state, while a graph neural scheduling layer represents crane interference, task precedence, berth adjacency, and yard-transfer coupling. A multi-objective decoder generates feasible rolling assignments that are evaluated in the digital twin before release. Experiments use a simulated coastal terminal containing 42 vessel calls, 9 quay cranes, 1,260 container task batches, and five nominal or disturbed operating scenarios. Relative to the rule-based baseline, the proposed model reduces mean vessel turnaround time by 18.7% and crane idle time by 21.4%. Under 20% sensor dropout, robust task completion improves by 15.9%, while normalized crane energy per handled container decreases by 9.8%. These results indicate that reliability-conditioned fusion, latent multimodal encoding, and graph-constrained candidate evaluation can jointly improve scheduling continuity under noisy and coupled terminal conditions. The findings support further field- oriented validation as a dispatcher decision-support method rather than immediate autonomous deployment.