Machine Learning for ETA Prediction

The significance of port operation efficiency has heightened with the increase in container volumes and vessel sizes. To enhance port operations efficiency, accurately predicting the estimated time of arrival (ETA) of sea-going vessels is crucial. Over recent years, numerous AIS data-driven methodologies have been proposed for estimating vessel ETA at ports. These techniques typically involve several steps. Initially, potential vessel trajectories are identified using AIS data mining methods, often employing reinforcement learning (RL). Subsequently, the speed over ground (SOG) of the vessels is estimated, utilizing various techniques such as Markov Chain properties and Bayesian sampling. Finally, the proposed method's performance is validated through experimental comparisons. Once the superiority of these methods is demonstrated, they can be applied in real-time scenarios to contribute to the development of an intelligent port system. Developing machine learning algorithms for vessel arrival time estimation poses a significant challenge for algorithm developers within the maritime sector.

The need for accurate estimation

Over 90% of global container shipments are transported by sea vessels. The global container port throughput experienced a 4.7% increase in 2018, a decline from the 6.7% rise observed in 2017, reaching a total of 793.26 million metric tonnes handled worldwide. The expansion of container traffic volume and vessel sizes are significant contributors to the exacerbation of port congestion. Approximately 93.6% of port delays are attributed to congestion. Addressing this issue necessitates the formulation of a logistics plan. However, a key challenge is the requirement for vessels to adhere to their schedules for such a plan to be viable. Unfortunately, the percentage of vessels maintaining their schedules falls within a rather discouraging range of 55-89%. Even when vessels do arrive punctually, there is a greater than 50% probability that they will follow vessels that were delayed by two to five days. The uncertainty surrounding vessel arrival times diminishes schedule reliability, leading to increased delays and decreased productivity levels for inland transport operators. Additionally, higher inventory levels are necessitated to meet service levels and mitigate disruptions to the production process, thereby escalating logistics costs. Vessel arrival delays also amplify the costs associated with vessel operation and supply chain management. For instance, a three-day delay due to port congestion and vessel waiting time incurs substantial round-trip costs. Adjusting berth allocations on the quay when a container vessel arrives late can lead to further delays as other vessels may need to reroute, heightening supply lead time variability. Frequent rescheduling of berth plans due to vessel delays at the terminal's first planning level compromises port efficiency. Efficient resource utilization necessitates a regular arrival pattern of vessels consistent with prior agreements or announcements by shipping companies when the port adjusts its berth plan. Hence, accurate prediction of estimated time of arrival (ETA) for vessels is crucial, as ports cannot devise efficient logistics plans in the absence of ETA certainty. ETA represents the anticipated date and time of a shipment's arrival at a specified destination.

Efforts on ETA prediction

Significant research efforts have been dedicated to enhancing the precision of Estimated Time of Arrival (ETA) prediction to support decision-making across various domains. In air traffic control, ETA prediction methodologies play a vital role in collaborative decision-making processes. One innovative approach involves employing a machine learning algorithm for ETA prediction. Similarly, in land transportation, studies have focused on accurately predicting ETA for buses and ambulances. Spatial analysis methods have been proposed for forecasting ETA of moving objects. Within the maritime sector, machine learning techniques, along with AIS data, are utilized to forecast the arrival of container ships. These studies introduce the application of predictive machine learning models to anticipate vessel routes and speeds.

Dependencies on AIS data

When predicting Estimated Time of Arrival (ETA) for a moving object, the trajectory of the object must initially be estimated. In maritime traffic, vessels' historical data is documented through Automatic Identification System (AIS) records. Mandated by the International Maritime Organization (IMO) in 2004, AIS facilitates the exchange of vessel data. As vessels navigate, they autonomously generate and log their historical trajectories in the AIS format. AIS data encompasses static, dynamic, and voyage-related information categories. Leveraging this data enables the prediction of a vessel's trajectory and Speed Over Ground (SOG). AIS data serves as a foundation for ETA prediction due to two key attributes of AIS systems: complete automation, eliminating the need for human intervention, and resilience to weather or sea conditions compared to radar due to longer wavelengths. This resilience enhances vessel movement monitoring, especially during nighttime. AIS systems are renowned for producing relatively accurate datasets, providing shipping companies with improved safety and operational benefits such as reduced collisions and pollution levels.

In modern ETA prediction algorithm development, several parameters are typically utilized. These include the Maritime Mobile Service Identity (MMSI), which serves as a unique identifier for vessels, vessel position, timestamp indicating when the signal was reported, and Speed Over Ground (SOG), representing the vessel's actual speed. MMSI distinguishes vessel trajectories, while the vessel's location and timestamp, linked to MMSI, contribute to trajectory prediction. ETA prediction relies on vessel speed, for which SOG, reflecting real speed, is a crucial factor.

Estimation of vessel arrival time via Reinforcement Learning

In recent years, numerous studies have endeavored to address the aforementioned challenges. However, many of these studies face limitations concerning trajectory prediction and the estimation of vessel speed. One prevalent approach in ETA prediction relies solely on historical vessel trajectory, yet it has constraints in considering various routes. To overcome this limitation, data-driven methodologies for vessel ETA prediction are gaining traction, leveraging innovative path-finding algorithms and speed over ground (SOG) estimation techniques. Reinforcement Learning (RL) frameworks are increasingly employed for predicting vessel trajectories, marking a departure from traditional methods. There is a pressing need for novel methodologies in vessel ETA prediction integrating artificial intelligence algorithms and resilient sampling methods tailored to AIS data challenges. A key distinction of these novel approaches from previous methods lies in their enhanced performance in congested areas with dense historical vessel trajectories.

Recent studies have introduced and applied novel methodologies to real-world scenarios using authentic datasets. These methodologies were compared with an alternative strategy involving traffic density information, directionality layers, and land-masking layers. Experimental findings indicate that the proposed methodologies enable RL agents to achieve higher accuracy in trajectory prediction. Moreover, the sampling methods employed for SOG estimation demonstrate superior accuracy and reduced variance. Additionally, the incorporation of artificial intelligence methods with historical data leads to greater performance improvements, particularly in heavily congested traffic areas.

Conclusion

Innovative approaches utilizing Reinforcement Learning (RL) for vessel arrival time estimation present opportunities for devising logistics plans for ports. Leveraging such systems can enhance port operational efficiency, thus alleviating port congestion. Despite the promising performance demonstrated by the proposed algorithm and methodology, there are inherent limitations, particularly in terms of results variability contingent upon RL configuration. To enable agents to identify trajectories over extended distances, future endeavors may explore the integration of long-distance path-finding algorithms employing RL techniques.