Adaptive Graph Attention and Long Short-Term Memory-Based Networks for Traffic Prediction

Zhu, Taomei; Boada, Maria Jesus Lopez; Boada, Beatriz Lopez · 2024 · Crossref

DOI: 10.3390/math12020255

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 traffic prediction, a critical component of intelligent transportation systems and smart city frameworks. While traditional statistical methods and early machine learning models have been widely used, recent advancements in deep learning have shifted the focus toward capturing complex spatial and temporal dependencies in traffic data. Existing hybrid models, such as Graph Convolutional Networks (GCNs) combined with Recurrent Neural Networks (RNNs), have improved performance but often struggle with decreasing accuracy over longer prediction horizons, sensitivity to hyperparameter tuning, and limited adaptability to dynamic traffic conditions or varying network structures. To address these limitations, the authors propose a novel hybrid architecture called GAT-LSTM, which integrates Graph Attention Networks (GAT) and Long Short-Term Memory (LSTM) networks. The proposed GAT-LSTM model is designed to jointly model graph-structured spatial features and sequential temporal dependencies. The architecture incorporates three main components: GAT layers to capture spatial relationships by assigning attention weights to neighboring nodes; LSTM layers to model temporal correlations and traffic patterns; and a "Dayfeature" component that encodes external factors such as time of day, day of the week, holidays, and extreme weather conditions. A key innovation is an adaptive attention block that dynamically learns and assigns weights to the outputs of the GAT network, the original traffic data, and the Dayfeature inputs before passing them to the LSTM network. This mechanism allows the model to adaptively combine these inputs for each specific sensor node, enabling local adaptation within the global model. The traffic network is represented as a dynamic graph where the adjacency matrix reflects connectivity and dynamic correlations between sensors. The model was evaluated using the PeMS08 open dataset, a standard benchmark for traffic prediction tasks. Experimental results demonstrate that the GAT-LSTM architecture achieves state-of-the-art performance in traffic flow prediction, outperforming existing baseline and hybrid models. The study highlights that the adaptive attention mechanism effectively improves prediction accuracy by accounting for varying spatial and temporal dependencies across different nodes. Furthermore, the model exhibits robust adaptability to dynamic traffic conditions, different prediction horizons, and various traffic network structures. The authors also note that the architecture facilitates the detection of weaker nodes within the traffic network, allowing for the design of local adaptation algorithms to further enhance performance. The significance of this work lies in its contribution to more flexible and accurate traffic prediction models. By integrating graph attention with LSTM and external feature encoding, the GAT-LSTM model addresses key challenges in handling non-Euclidean spatial structures and long-term temporal dependencies. The adaptive nature of the model reduces the reliance on extensive hyperparameter tuning for specific datasets and enhances generalizability across different traffic networks. This approach provides a robust framework for improving traffic management, congestion reduction, and transportation planning in smart city environments.

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