关键词: 集装箱码头, 集卡周转时间, 极端梯度提升算法, 贝叶斯优化

Abstract: A truck's turn time in a port is from the period when it enters the gate, with all loading and unloading tasks completed in the designated blocks, to the period when the truck exits the gate. Inland transit and container terminals both rely heavily on trucks. A precise long-term forecast of the average turn time of the truck can help the truck company or logistics provider plan their fleet of vehicles to avoid arriving at peak time, which not only cuts down on turn time in the truck port but also speeds up the turnover rate for collection and distribution. It can also improve the effectiveness of truck dispatching and lower overall business costs. The following points are the primary contributions of this paper: (1)This paper establishes a prediction model for the following 24 hours of truck turn time, different from short-term prediction, which gives truck companies a foundation to improve scheduling. (2)To accomplish accurate prediction of large-scale data, the best combination of hyperparameters is found based on XGBoost Tree (eXtreme Gradient Boosting-decision Tree) and Bayesian optimization (BO). The benefits of this model are examined from the theoretical principles and data findings viewpoints. (3)The forecast is more accurate because of input features such as internal workload, service types, and historical turn time.
Firstly, this paper proposes a long-term prediction method for truck turn time, based on the XGBoost-Tree. To speed up the convergence of training, Bayesian optimization is used to search the global optimized combination of hyperparameters. Grid search is a time-consuming method for finding combined parameters. The search effect of Bayesian optimization is quite similar to grid search, however, it has a far higher search efficiency. While random search is quick, its effectiveness is inferior to that of Bayesian optimization.
Secondly, this paper makes use of approximately 5.797 million records of pertinent operation data from a Shenzhen port from September 1, 2018, to August 31, 2019. It provides the relevant ships' arrival, berthing, and departure times as well as the external truck's entrance, completion, and departure times, service types, and container classification. New features, which have not been used in previous papers, are added to the XGBoost-Tree to predict the turn time of t hour. They are truck turn time in (t-24) hour, service type, and internal workload in the hour. In contrast to a short-term forecast, which predicts the turn time of the following period, a long-term forecast creates a prediction model for the turn time of the following 24 hours, which matches the schedule horizon of the truck company. Additionally, it is found that the 24-hour turn time prediction relates to the autoregressive curve of turn time significantly. The turn time of various operation types each has its unique characteristics. The turn time histogram of the four different operation types reveals that the external truck of the dual tasks has the longest turn time because it involves unloading and loading in two separate blocks. Besides, the full containers for import will encounter the inevitable turnover operation, so the turn time is also lengthy. The internal workload in the port is a significant factor that influences the turn time of the trucks. It can be seen from the scatter chart of the internal workload and average turn time of trucks. This is because the trucks must wait in line when the operation equipment is busy.
Thirdly, the results show that XGBoost-Tree outperforms XGBoost-Linear, Random Forest, Support Vector Regression, Recurrent Neural Network, and Long-Short Term Memory, with an improved accuracy rate by 7.0% (Empty Container for Import), 0.8% (Full Container for Import), 13.44% (Full Container for Export) and 9.8% (Dual-Tasks). The statistical approach is also contrasted with the model in this work. According to the statistical approach, there is a linear or binary relation between the turn time and the truck's arrival. The comparative analysis demonstrates that the machine learning approach has a greater level of fitting accuracy. Service types, as a combination of human experience and machine intelligence, improve accuracy rates of 11.1% (Empty Container for Import), 8.8% (Full Container for Import), 26.4% (Full Container for Export), and 19.6% (Dual-Tasks). And, this model takes into account the nonlinear effect on the turn time from service types. The prediction error is less than those based on separate operation types.
Finally, the sensitivity analysis demonstrates that interactions of turn time among the hour, service type, and volumes of transactions in the hour are of great significance in the prediction models. The future work will consider more input features related to the turn time of trucks, and deeply discuss the positive impact of the release of turn time prediction on the decision-making of the truck companies.

Key words: container terminal, truck turn time, XGBoost algorithm, Bayesian optimization