DeepST-Net: Leveraging Spatiotemporal Data Refinement for Enhanced Travel Time Prediction

Abstract

Accurate travel time prediction is essential for optimizing transportation systems and alleviating congestion in urban environments. This study introduces a comprehensive framework that combines advanced spatiotemporal data refinement techniques with a deep learning model, DeepST-Net, to improve prediction accuracy. Using a dataset from the New York City Taxi and Limousine Commission (TLC), an ablation study is conducted to systematically assess the impact of various preprocessing strategies on model performance. Additionally, the DeepST-Net model is benchmarked against established models, including MURAT, ST-NN, XGB-IN, and JSTC. Our findings demonstrate that DeepST-Net, leveraging refined spatiotemporal features, achieves significant improvements in prediction accuracy, with a mean absolute error (MAE) of 59.27 s, a root mean square error (RMSE) of 89.63 s, and a mean absolute percentage error (MAPE) of 9.73%, outperforming all benchmark models. By bridging data refinement and predictive modeling, the proposed framework offers a scalable and effective solution for travel time prediction in complex urban contexts.

Keywords

• travel time prediction
• spatiotemporal data
• deep learning
• transportation systems