A Hybrid Deep Learning Network for Long-Term Travel Time Prediction in Freeways

Abstract

Travel time plays a vital role in people’s daily lives. It can help them not merely avoid traffic congestion but save time as well. When people need to drive to different cities by taking highways, travel time become more and more important now that they can check it to arrange better routes. Moreover, because COVID-19 are epidemic across Taiwan, people prefer to drive rather than taking public transportation. Therefore, accurate predictions of travel time is of great significance. In order to obtain precise predictions and correspond to situations in real life, we divide data into long and short sequences and create three types of dataset, including the whole year, only national holidays, and non-holidays. Additionally, on account of the interactive influence of time in different segments of the freeway, we exploit data to predict next-hour travel time instead of next 5 minutes. We introduce a deep learning model which hybrids tendency from XGBoost and recency embeddings from a fully-connected neural network, respectively. It can capture crucial features of both long and short sequences and observe implicit correlations between XGBoost and a fully-connected neural network. Extensive experiments on the dataset illustrate that our model achieves eminent performances and outperforms other state-of-the-art models.