Periodic Attention-based Stacked Sequence to Sequence framework for long-term travel time prediction

Abstract

Travel time analysis and prediction are keystones for building intelligent transportation systems in the new era, which has gained wide attention from the research community. Over the last few years, deep learning-based approaches have demonstrated the capability to predict travel time with big data. However, the existing works focused on predicting short-term travel time for paths and roads. In fact, knowing a period of long-term future travel time is a demand with important applications like traffic management and schedule routing planning. Nevertheless, studies on long-term traffic prediction are still very limited due to the complicated underlying factors. In this paper, we propose a novel deep learning-based framework named Periodic Attention-based Stacked Sequence to Sequence (PASS2S), which aims to address the long-term traffic prediction problem. PASS2S consists of two main components, namely periodic segment generation and attention-based stacked sequential prediction. To extract periodic information, we design a periodic segment generation component to capture historical periodic segments from the traffic data. To reduce the error propagation and improve long-term prediction accuracy, we propose an attention-based stacked prediction component to model long short-term and short-term dependencies from the periodic segments. We conducted a series of experiments on a real-world travel time dataset and the experimental results show that our proposed approach outperforms the state-of-the-art competing methods in terms of various metrics like MAE, RMSE, and SMAPE. To the best of our knowledge, this is the first work that considers attentive periodic historical information for solving the long-term travel time prediction with a period of the future, which has not been well studied in the research community.