Accurate Estimated Time of Arrival (ETA) forecasting is essential for improving operational planning and decision-making in modern ports. While machine learning has significantly enhanced ETA prediction using Automatic Identification System data, the impact of data preprocessing on forecasting performance remains underexplored. This study investigates how AIS data preprocessing and machine learning model selection jointly affect ETA forecasting for short-sea shipping, using the Port of Sines as an empirical case study. A reproducible forecasting workflow was developed, integrating dataset construction, voyage selection, feature engineering, predictive modelling and performance evaluation. Three supervised machine learning algorithms, K-Nearest Neighbors, Random Forest Regression, and Multilayer Perceptron, were trained and compared under identical experimental conditions. The results show that forecasting performance depends not only on model selection but also on the quality of the modelling dataset. In particular, Random Forest Regression achieved the strongest and most consistent performance. It was found to be invariant to feature scaling, whereas scaling had a small negative effect on K-Nearest Neighbors and increased training variance for the Multilayer Perceptron. Although all three models achieved accurate ETA predictions, they exhibited different strengths regarding predictive performance, computational efficiency, and operational applicability. The proposed workflow contributes to the development of transparent and reproducible ETA forecasting methodologies and provides practical guidance for implementing AIS-based decision-support systems in short-sea port operations.
Big Data platforms are expected to unify data engineering, governance and machine learning in one environment, and the data lakehouse has become the dominant architectural answer. Empirical evaluations on real operational data remain scarce, and those that exist tend to validate a downstream model rather than the architecture itself. This study designs, implements and evaluates a governed data lakehouse following Design Science Research. The artefact organises Apache Spark, Delta Lake, Unity Catalog and MLflow into a medallion architecture, and is exercised on operational data from the Port of Sines: 36,773 reporting events yielding 8148 berth visits by 1923 vessels between October 2020 and September 2024. The architecture is evaluated against explicit criteria rather than by proxy. Ingestion throughput stays between 16.8 and 18.5 thousand records per second as volume quadruples; the curated Gold layer answers analytical queries 100 times faster than recomputation from source and compacts storage 6.6-fold; an explicit quality ledger and a physical-bound quarantine make governance auditable and measurably beneficial downstream, restoring a correlation structure that 55 out-of-bound values had erased; and an incremental refresh path updates the analytical layer in 0.98 ms at the weekly cadence, at a cost that does not grow with accumulated history. A berth-time regression under rolling-origin validation, benchmarked against tuned ensembles, operational heuristics and time-aware sequence models, serves as the analytical probe rather than as the contribution. The study contributes a measured, replicable reference architecture for governed Big Data analytics.
Background: Accurate prediction of vessel time at berth is essential for smart berth planning, yet berth-allocation research typically treats processing time as a given input rather than a learned variable. This study develops a governed data lakehouse blueprint for port operational data and uses berth-time prediction at the Port of Sines, Portugal, to establish what routinely collected administrative records support. Methods: Following a design science research approach, a Bronze–Silver–Gold medallion pipeline was realised twice from one design, on a managed cloud platform and on single-node open-source components, demonstrating portability rather than vendor dependence. Four years of administrative data (October 2020 to September 2024) yielded 8148 berth visits. Results: Under rolling-origin temporal validation across five cut-offs, a Random Forest regressor attained a mean R2 of 0.260 and a mean absolute error of 10.1 h, against 13.5 h for the median-stay heuristic, predicting 52.6% of visits within six hours. SHapley Additive exPlanations identified the vessel’s own previous stay, the berth and vessel size as the strongest predictors. Pipeline measurement shows near-linear scaling and a hundredfold query speed-up over raw-source recomputation. Conclusions: Predictive accuracy remains bounded by the absence of exogenous operational variables; the contribution is the governed data foundation that berth planning requires.