Comparison of short-term traffic demand prediction methods for transport services

Abstract

The purpose of this paper is to compare the short-term traffic demand prediction methods for transport services. Forecasting traffic demand throughout a city can help to organize the vehicles and minimize the waiting time for passengers and drivers. In addition, the prediction result is an important resource of other studies such as dispatching systems and recommendation systems. Therefore, short-term demand prediction is significant. Several methods for short-term traffic demand have been developed recently. However, the existing literature lacks studies that focus on how to choose the appropriate prediction method. This study implements six time-series models, three machine learning models, and three deep learning models on more than one datasets. The results reveal that no one model could get the best performance across all times, all areas or different cities. Random forest and XGBoost could get higher performance with a short time. Deep learning methods are efficient in other areas, but they do not necessarily perform better than other methods. Embedding is a good way to save time when calibrating a model compared with one-hot encoding. One cannot depend on increasing the size of the training data set to obtain the best performance.