Container terminals are vital hubs in global trade, facilitating the seamless transfer of goods between maritime and inland transportation networks. Truck turnaround time serves as a critical performance metric for container terminals, provides direct feedback on port congestion and efficiency. This study focuses on developing a predictive framework for truck turnaround time (TTT) at the Port of Rotterdam by integrating multi-source data and employing advanced machine learning techniques.
Previous studies on truck turnaround time prediction have largely relied on limited datasets and methodologies, such as utilizing historical truck arrival flows or terminal operation logs and using statistical methods. This research employs diverse datasets, including Bluetooth detection records, container arrival information, and environmental condition. By combining these data sources, a harmonized dataset was constructed to represent the complexities of port operations. A stacked Long Short-Term Memory (LSTM) network was employed as the predictive model, utilizing its ability to capture temporal dependencies and nonlinear interactions between variables. This approach allows for more comprehensive and accurate TTT predictions compared to conventional methods.
To process the noisy and incomplete Bluetooth data, a robust trip identification pipeline was developed. The pipeline employed spatial clustering, temporal filtering, and dual verification to accurately identify container truck trips, achieving an accuracy exceeding 90%. Using this processed data, the stacked LSTM model demonstrated superior predictive performance, effectively capturing periodic trends and long-term dependencies. Benchmarking results showed that the stacked LSTM outperformed traditional methods, including Random Forest and XGBoost. Sensitivity analysis highlighted the critical role of truck arrival flows and wind conditions in determining truck turnaround time variability.
In summary, this study provides a novel and scalable framework for TTT prediction, integrating multi-source data and advanced modeling techniques to address key limitations of existing approaches. The findings offer actionable insights for optimizing terminal operations and reducing congestion. Future research could focus on expanding data sources, enhancing model interpretability, and validating the framework across diverse port environments to ensure broader applicability.