T-LSTM: A Long Short-Term Memory Neural Network Enhanced by Temporal Information for Traffic Flow Prediction

Mou, Luntian; Zhao, Pengfei; Xie, Haitao; Chen, Yanyan · 2019 · OpenAlex-citations

DOI: 10.1109/access.2019.2929692

archive: archived pipeline: cataloged verified

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This paper addresses the challenge of accurate short-term traffic flow prediction, a critical component of Intelligent Transportation Systems (ITS) for traffic guidance and congestion mitigation. While Long Short-Term Memory (LSTM) networks have demonstrated strong performance in this domain, existing methods often neglect the intrinsic correlation between traffic flow and temporal information, such as the recurring daily patterns of commuting hours. The authors argue that incorporating explicit temporal features can enhance the model's ability to capture time-variant fluctuations. Additionally, the paper tackles the issue of missing data in traffic sensor records, proposing a novel repair technique to improve prediction accuracy. To address these gaps, the authors propose a Temporal information enhancing LSTM (T-LSTM) model. This architecture integrates recurrent time labels with the standard LSTM structure, allowing the network to learn higher-level temporal representations. The input vector consists of both the traffic flow value and a corresponding time label, enabling the model to recognize periodic trends. The study utilizes real-world traffic detector data from the East Fourth Ring Road in Beijing, collected from March to August 2014. The data, originally sampled every two minutes, was aggregated into 16-minute intervals. The dataset was split into training (March–July) and testing (August) sets, with normalization applied. The model was implemented using TensorFlow and Python, featuring three stacked LSTM layers. Hyperparameters were optimized using the Adam optimizer and Mean Square Error loss function. Experimental results demonstrate that T-LSTM outperforms several state-of-the-art models, including standard LSTM, Gated Recurrent Unit (GRU), Stacked Autoencoder (SAE), Deep Belief Network (DBN), Support Vector Machine (SVM), K-nearest neighbor (KNN), Feed Forward Neural Networks (FFNN), and ARIMA. Specifically, T-LSTM achieved a Mean Absolute Percentage Error (MAPE) of 6.09% on the test set. Compared to a standard LSTM without time labels, T-LSTM reduced RMSE by 13.4% and MAPE by 1.44%. The study also validated a T-LSTM-based missing data repair technique, which effectively restored the characteristics of original data by inferring missing values from historical sequences, thereby further improving prediction accuracy compared to simple historical averaging methods. The significance of this work lies in demonstrating that explicit temporal information is a critical feature for enhancing short-term traffic flow prediction accuracy. By integrating time labels into the input structure, T-LSTM captures periodic traffic patterns more effectively than models relying solely on flow data. Furthermore, the proposed missing data repair method offers a robust solution for handling incomplete sensor data, a common issue in ITS applications. These findings suggest that future traffic prediction models should prioritize the integration of temporal features and advanced data preprocessing techniques to achieve higher reliability and precision.

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed.

Summary generated by qwen3.6-27b-prismaquant on 2026-06-26; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• anticipation