Interpretable Machine-Learning Warning of Tropical-Cyclone-Induced Port Service Disruption Using Automatic Identification System and Forecast Advisories

Using six US hurricane-strength tropical-cyclone events as the independent transfer units, this study evaluates advisory-time warnings of port service disruption from public automatic identification system (AIS) activity, National Hurricane Center advisories, and port-network data. The 19-port panel contains 3312 repeated advisory-lead-port rows, but leave-one-event-out validation is limited by the six-event ceiling. For a 10% AIS-based service-loss label, the full calibrated logistic model has a mean area under the receiver operating characteristic curve (AUROC; where 0.5 denotes random ranking) of 0.702 (95% event-bootstrap interval: [0.487, 0.864]) and an F1 score (the harmonic mean of precision and recall) of 0.540 (95% interval: [0.370, 0.752]). Thus, absolute performance is uncertain. The more robust comparative result is that forecast exposure alone has AUROC = 0.427, below random ranking, whereas adding a recent port state and network substitutability improves the ranking. Several simple rules match or exceed the model’s F1 score only by warning almost every sample; calibrated probabilities instead support an explicit precision-recall-warning-burden tradeoff. Sally 2020 is nearly separable because static network variables distinguish four positive ports from two negative ones, whereas Ida 2021 fails through a corridor-level ranking reversal. The framework is, therefore, a promising, interpretable decision-support prototype, not evidence of universal predictive performance.