Can AIS data improve the short-term forecast of weekly dry bulk cargo port throughput? - a machine-learning approach
Minato Nakashima, Ryuichi Shibasaki
Abstract
This study examines the development of a machine-learning model to forecast weekly throughputs of dry bulk cargo in the short term based on automatic identification system (AIS) data. Specifically, the weekly amounts of iron ore exported from several major ports in Australia and Brazil in the latter half of 2019 are forecasted three weeks in advance using a long short-term memory model. We examine many variables extracted from AIS data, including the vessel position, speed, draught, and destination, as the input features of the model. Consequently, we develop a highly accurate forecasting model that uses four influential variables derived from AIS data, namely, vessel traffic around the target port and in the region, vessel traffic at major partner import ports, and vessel traffic at the target port during the past year. Finally, by forecasting the weekly port cargo throughputs in the first half of 2020, which was affected by the COVID-19 pandemic, the applicability of the model is confirmed, even for ports where the throughput fluctuates significantly. In particular, this study demonstrates that AIS data are beneficial not only as a real-time traffic database but also as a database containing various related explanatory variables, including historical vessel traffic.