Machine Learning Approach to Forecasting Empty Container Volumes

• Main Author: LIU, YUAN
• Format: Others
• Language: English
• Published: Blekinge Tekniska Högskola, Institutionen för datavetenskap 2019
• Subjects: Machine learning; Regression; Empty Containers; Forecasting; Time Series; Computer Sciences; Datavetenskap (datalogi)
• Online Access: http://urn.kb.se/resolve?urn=urn:nbn:se:bth-18858

Background With the development of global trade, the volume of goods transported around the world is growing. And over 90% of world trade is carried by shipping industry, container shipping is the most important way. But with the growth of trade imbalances, the reposition of empty containers has become an important issue for shipping. Accurately predicting the volume of empty containers will greatly assist the empty container reposition plan. Objectives The main aim of this study is to explore the effect of machine learning in predicting empty container volumes, make a performance comparison and analysis with existing empirical methods and mathematical statistics methods. Methods The main method of this study is experiment. In this study I chose the appropriate algorithm model and then trained and tested the model. This study uses the same data sources as the industrial approach, using the same metric to evaluate and compare the performance of machine learning methods and industrial methods.  Results Through experiments, this study obtained the forecasting performance results of five machine algorithms including the LASSO regression algorithm on the Los Angeles Port and Long Beach Port datasets. Metrics are (Mean Square Error) MSE and (Mean Absolute Error) MAE. Conclusions LASSO Regression and Ridge Regression are the best machine learning algorithms for predicting the volume of empty containers. Compared to empirical methods, the single machine learning algorithm performs better and has better accuracy. However, compared with mature statistical methods such as time series, the performance of a single machine learning algorithm is worse than the time series method. Machine learning needs to try to combine multiple models or select more high-correlation feature quantities to improve performance on this prediction problem.