DAT-STAN - Dual-Module Adaptive Transformer and Spatio-Temporal Attention Network for Large-Scale Traffic Flow Prediction

Sep 17, 2026

Traffic flow prediction is fundamental to urban traffic management and congestion prevention within intelligent transportation systems, especially as large-scale real-time traffic data becomes widely available. While graph neural networks and Transformer-based models have achieved strong results, many existing approaches do not fully model evolving traffic patterns and complex node interactions over time, nor do they effectively fuse spatio-temporal and multi-angle features for reliable long-horizon forecasting. This paper introduces DAT-STAN to address these issues through dynamic, data-driven structures and spatio-temporal attention.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages17

Reading Time43 min

AccessFree to download

Introduction

Related work and challenges

Proposed approach: DAT-STAN

Experiments and results

TSINGHUA SCIENCE AND TECHNOLOGY ISSN 1007-0214 04/37 pp45 −61 DOI: 10 . 26599/TST . 2025 . 9010062 Volume 32, Number 1, February 2027 Research Article DAT-STAN: Dual-Module Adaptive Transformer and Spatio-Temporal Attention Network for Large-Scale Traffic Flow Prediction Kuanye Yin, Xinzheng Niu* , Jiahui Zhu, Yuemei Jiang, and Fan Min Abstract: Traffic flow prediction is crucial for managing urban traffic and preventing congestion in intelligent transportation systems, particularly with the rise of large-scale real-time traffic data. Graph neural networks and Transformer-based architectures have shown significant success in this task. However, existing methods struggle to: (1) fully capture the evolving traffic patterns and complex interactions between nodes over time, and (2) effectively integrate spatio-temporal and multi-angle features, leading to less accurate long-term predictions. In this paper, we propose the Dual-module Adaptive Transformer and Spatio-Temporal Attention Network (DAT-STAN) to address these challenges in the context of large-scale traffic data. The model utilizes dynamic, data-driven structures to capture evolving relationships in the traffic network and incorporates spatiotemporal attention mechanisms to focus on key features while reducing irrelevant information. Extensive experiments on multiple large-scale, real-world datasets show that DAT-STAN outperforms state-of-the-art models in both prediction accuracy and computational efficiency. Key words: traffic flow prediction ; spatio-temporal attention ; Transformer; adaptive adjacency matrix ; dynamic traffic conditions 1 Introduction With the advancement of transportation systems and the proliferation of urban intelligence, vast amounts of real-time traffic data are now being generated across urban networks. This surge in data, coupled with the increase in vehicle ownership, has made big data- Kuanye Yin, Xinzheng Niu, Jiahui Zhu, and Yuemei Jiang are with School of Computer Science and Engineering, University of Electronic Science and Technology of China, Chengdu 611731, China. E-mail: [202321080614@std.uestc.edu.cn](202321080614@std.uestc.edu.cn); [xinzhengniu@uestc.edu.cn](xinzhengniu@uestc.edu.cn); [jiahui.z@qq.com](jiahui.z@qq.com); jisuanjym@ [126.com](126.com). Fan Min is with School of Computer Science and Software Engineering and Laboratory of Machine Learning, and Institute for Artificial Intelligence, Southwest Petroleum University, Chengdu 610500, [China. E-mail: minfan@swpu.edu.cn](China. E-mail: minfan@swpu.edu.cn). * To whom correspondence should be addressed. Manuscript received: 2024-11-24; revised: 2025-02-19; accepted: 2025-04-10 driven traffic flow prediction within Intelligent Transportation Systems (ITS) a critical research focus[1] . However, the gap between limited computing resources and the rapid growth of data necessitates low-consumption approaches for efficient traffic flow prediction. Accurate traffic flow prediction leveraging big data analytics is essential for enabling informed urban planning, alleviating congestion, and optimizing traffic signal timings, which collectively enhance the efficiency of road networks[2] . Consequently, substantial research efforts and resources have been dedicated to this domain, leveraging advanced methodologies, such as machine learning, big data analytics, and real-time data processing. Early efforts in traffic flow prediction mainly utilized time series analysis models, such as the AutoRegressive Integrated Moving Average (ARIMA)[3] and the Vector AutoRegressive (VAR)[4] . While these methods effectively extract temporal features, capturing the non- © The author(s) 2027. The articles published in this open access journal are distributed under the terms of the Creative Commons Attribution 4.0 International License ([http://creativecommons.org/licenses/by/4.0/](http://creativecommons.org/licenses/by/4.0/)). 46 Tsinghua Science and Technology, February 2

FAQs

What problem does DAT-STAN target in traffic flow prediction?

It targets the difficulty of capturing evolving traffic patterns and complex node interactions over time, and the challenge of effectively integrating spatio-temporal and multi-angle features for more accurate long-term forecasts.

How does DAT-STAN address spatio-temporal modeling?

It uses dynamic, data-driven structures to capture evolving relationships in the traffic network and employs spatio-temporal attention mechanisms to focus on key features while reducing irrelevant information.

What evidence is provided for DAT-STAN’s effectiveness?

Extensive experiments on multiple large-scale, real-world datasets show DAT-STAN outperforms state-of-the-art models in both prediction accuracy and computational efficiency.

Sep 17, 2026

Uploaded by 8796093062539