Kernel-Based Machine Learning Models for Predicting Daily Truck Volume at Seaport Terminals

Abstract

The heavy truck traffic generated by major seaports can have huge impacts on local and regional transportation networks. Both transportation agencies and port authorities have a need to know in advance the amount of truck traffic in order to accommodate them accordingly. Several previous studies have developed models for predicting the daily truck traffic at seaport terminals using terminal operation data. In this study, two kernel-based supervised machine learning methods are introduced for the same purpose: Gaussian processes (GPs) and e-support vector machines (e-SVMs). They are compared against the multilayer feedforward neural network (MLFNN) model, which was used in past studies, to provide a comparison of their relative performance. The model development is done using the data from Bayport and Barbours Cut container terminals at the Port of Houston. Truck trips generated by import and export activities at the two terminals are investigated separately, generating four sets of data for model testing and comparison. For all test data sets, the GP and e-SVM models perform equally well and their prediction performance compares favorably to that of the MLFNN model. On a practical note, the GP and e-SVM models require less effort in model fitting compared to the MLFNN model. The strong performance of the GP and e-SVM models and their relative ease of use make them viable alternative approaches to the MLFNN in port-generated truck traffic predictions.