Spatio Temporal-Graph Convolutional Networks를 활용한 국내 항만 교통량 예측

Abstract

Ports are located at the intersection of the sea and inland regions, facilitating the flow of logistics between countries and playing an essential role in the country economic development in terms of manufacturing, trade, fishing, energy transport, and tourism. Ports tend to increase the Gross Domestic Product(GDP) of the region where the port is located and have a large positive ripple effect on the GDP of the surrounding region. Due to these economic advantages, competition in ports around the world is getting fiercer. Modern port competitiveness, which is fierce around the world, is not evaluated only by reducing the waiting time of incoming and outgoing ships by improving work efficiency, service quality, and productivity. External parts of the traditional port logistics industry, such as port air pollutant control and port safety accident management, which have recently become increasingly important, are evaluated as port competitiveness. From this point of view, the concept of a smart port that improves port productivity and realizes eco-friendly port operation at the same time based on digital information technology is attracting attention. For the successful strategic promotion of smart ports, port traffic forecasting is essential. In the past, traditional methods such as simulation and statistical models were used to predict port traffic volume. However, recently, time series prediction models using deep learning are attracting attention due to the rapid change in the shipping and port logistics industry. However, Recurrent Neural Networks(RNN), a representative time-series prediction deep learning model, is a model specialized for sequence-type data and reflects temporal characteristics but has limitations in not considering spatial correlation, and Convolutional Neural Networks(CNN) reflect spatial features but do not reflect actual connection features. A port is not operated as a single entity but is operated as a network composed of countless connections between other ports. Therefore, to analyze the temporal characteristics of the port space, it is necessary to analyze the data in the form of an organic network. Thus, this study constructed a network of Korean ports using the Automatic Identification System(AIS) dataset containing vessel operation information and the spatial information data of Korean ports. Using Graph Convolutional Networks(GCN) based Spatio Temporal Graph Convolutional Networks(ST-GCN) model, I performed the port traffic prediction considering the spatial features connected between ports. This study utilized comparative models such as RNN-type Long Short-Term Memory(LSTM) and CNN-type 1D-CNN. Port entry and departure prediction results showed the ST-GCN performance improvement of about 5% over LSTM and about 7% over 1D-CNN. The study results are expected to provide insight for port authorities, logistics companies, and port operators to establish effective policies such as port management, cost-effective route development, port transportation, and systematic and efficient smart port construction.