A Data-Driven Decision-Making Framework for Dynamic Pilot Transfer Point Optimisation in Port Approaches

Camliyurt G., Turgut B. Ş., Park H., Kim S. H., Kim D., Park Y., ...Daha Fazla

Reliability Engineering & System Safety, ss.113821, 2026 (Scopus)

Özet

Maritime pilot transfer is among the most safety-critical manoeuvres in port approaches. This study presents a data-driven framework that predicts where a pilot transfer will take place when the pilot boat is dispatched, using only information available at that moment, and relocates the predicted point where hard safety constraints or a clearly better safety score require it. The framework integrates Automatic Identification System (AIS) trajectories, electronic navigational charts, tide-adjusted under-keel clearance (UKC), metocean observations and traffic context. Applied to one year at Busan Newport, it confirms 3,243 transfers by the 13 port pilot boats against the official pilotage log. At a dispatch proxy identified retrospectively from AIS, inbound meeting distances are predicted with R2=0.495 (mean absolute error, MAE = 1.28 km) under ship-grouped cross-validation and 0.537 in a forward temporal hold-out (0.30 and 0.32 without the transfer-time bound on the dispatch search); outbound distances are not predictable from the Busan AIS, which excludes the berths. The official fixed boarding zones perform far worse than any data-driven model. An out-of-fold optimisation using only AIS up to dispatch and chart-based routes finds a feasible recommendation in 96.4% of 1,955 cases, all with UKC ≥ 2 m and 83.1% in the low collision-risk class, with a median relocation of 300 m. At 11 further stations in the United States and Denmark, locally retrained models reach inbound R2 of 0.50–0.92 at ten stations. A size-corrected closest-point-of-approach (CPA) metric is checked against oriented hull polygons. The results support treating nominal boarding zones as dynamic corridors planned at pilot-boat dispatch.