Predicting Maritime Network Links for Resilient Strategic-Commodity Supply Chains: A Dynamic Edge-Weighted GCN–LSTM with Multidimensional Disturbance Factors

Maintaining reliable maritime connectivity is essential to resilient strategic-commodity supply chains, yet link dynamics and external disturbances are often modeled separately. This study develops a dynamic edge-weighted graph convolutional network–long short-term memory (GCN–LSTM) for one-month-ahead prediction of country-to-country maritime links. Monthly origin–destination connectivity among 115 countries or regions was analyzed for coal, iron ore, crude oil, liquefied natural gas, liquefied petroleum gas, and soybean from April 2014 to May 2024. Node features combine directed degree measures with 15 environmental, socioeconomic, and security factors; link-level trend, seasonality, and multiscale periodicity determine dynamic edge weights. A chronological training–validation–test split was used. Across the six networks, the full model achieved mean precision, recall, F1, and PR-AUC values of 0.6165, 0.6436, 0.6290, and 0.6626. Within the ablation study, the disturbance-only configuration produced a slightly higher mean PR-AUC (0.6681), whereas the full model achieved the highest mean F1. In the five-seed architecture benchmark, the proposed model achieved a cross-commodity mean F1 of 0.6258 ± 0.0048 and a mean PR-AUC of 0.6648 ± 0.0090 across five runs, exceeding GCN, Informer, and Crossformer but remaining below the standalone LSTM and Transformer baselines. Replacing dynamic edge weights with unit weights reduced mean F1 and PR-AUC by 10.6% and 13.7%. Dynamic edge weighting produced the most consistent architecture-related performance gain across the six networks. The relative predictive utility of the disturbance and link-evolution feature groups varied across commodities; the present analysis did not estimate the predictive importance of individual disturbance variables. Month-ahead link probabilities can serve as screening signals for prioritizing country pairs for further analysis.