Multi‐Step Ahead Traffic State Prediction Using a Stochastic Residual Long Short‐Term Memory and Temporal Dilated Attention Mechanism
IET Intelligent Transport Systems · 2026-01 · 20 (1)
DOI: 10.1049/itr2.70243 ↗Kurz gefasst
Fachartikel in der Zeitschrift IET Intelligent Transport Systems, online erschienen am 21.08.2026. Verfasst von Selim Reza und 3 weiteren Autor:innen. Die Arbeit nennt 79 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY-NC-ND 4.0). Gefördert von: European Regional Development Fund.
Abstract (Originalsprache)
Accurate multi‐step prediction is crucial for anticipating future traffic conditions, but existing models frequently encounter performance degradation due to external noise and error accumulation over longer prediction horizons. This study presents a multi‐step‐ahead traffic state prediction model that combines stochastic residual long short‐term memory (SResLSTM) networks with a temporal dilated attention (TDA) mechanism to address these challenges. The SResLSTM architecture mitigates vanishing or exploding gradient issues, facilitates learning long‐term temporal dependencies, and improves network training efficiency. It also incorporates Gaussian noise with the inputs to enhance the model's robustness. The TDA mechanism comprises multi‐scale causal convolutions, multi‐head attention, and gated residual fusion to enhance the model's prediction accuracy by dynamically assigning relevant weights to different time steps, thereby capturing the temporal variability essential for precise multi‐step prediction. This research also proposes a Fourier denoise filter (FDF) using a learnable frequency‐domain denoising technique to suppress high‐frequency noise from the input data. The proposed model was evaluated using three benchmark traffic flow and speed datasets: the PEMS, PEMS‐BAY, and METR‐LA datasets. Extensive experiments demonstrated good performance improvements across different prediction lengths on all three datasets compared to the baselines, such as a 6.97%, 10.16%, and 0.55% enhancement in terms of mean absolute error, root mean squared error, and score, respectively, for a prediction horizon of 60 min based on the PEMS‐BAY dataset. It also demonstrated enhanced robustness to rare events, random sensor dropouts, and adversarial attacks of various magnitudes. However, the model requires moderate computational resources, with a training time of 491.22 s on the PEMS dataset and a larger convergence time than the baselines. Additionally, the good outcomes of the conducted generalizability test on the European road freight transport flow dataset further verified the proposed model's worthiness.
Förderung
- European Regional Development Fund
Zitieren
APA
Reza, S., Ferreira, M. C., Machado, J. J. M., & Tavares, J. M. R. S. (2026). Multi‐Step Ahead Traffic State Prediction Using a Stochastic Residual Long Short‐Term Memory and Temporal Dilated Attention Mechanism. IET Intelligent Transport Systems, 20(1). https://doi.org/10.1049/itr2.70243
Vancouver
Reza S, Ferreira MC, Machado JJM, Tavares JMRS. Multi‐Step Ahead Traffic State Prediction Using a Stochastic Residual Long Short‐Term Memory and Temporal Dilated Attention Mechanism. IET Intelligent Transport Systems. 2026;20(1). doi:10.1049/itr2.70243
BibTeX
@article{Reza2026,
author = {Reza, Selim and Ferreira, Marta Campos and Machado, José J. M. and Tavares, João Manuel R. S.},
title = {Multi‐Step Ahead Traffic State Prediction Using a Stochastic Residual Long Short‐Term Memory and Temporal Dilated Attention Mechanism},
journal = {IET Intelligent Transport Systems},
year = {2026},
volume = {20},
number = {1},
doi = {10.1049/itr2.70243}
}
EUVIMED