A spatiotemporal graph Transformer framework that jointly models spatial interactions and temporal dependencies for traffic forecasting in edge computing and leverages Transformer-based self-attention to learn long-range temporal patterns from historical traffic observations is proposed.
Abstract
Accurate traffic forecasting is essential for proactive resource management in edge computing, where service demand evolves dynamically across both space and time. In practical cellular edge systems, traffic exhibits strong spatial correlations among neighboring service regions and long-range temporal dependencies driven by user mobility and application behavior. Existing recurrent forecasting approaches can capture short-term dynamics but often struggle to model long-horizon traffic evolution under non-stationary conditions. To address this challenge, we propose a spatiotemporal graph Transformer framework that jointly models spatial interactions and temporal dependencies for traffic forecasting in edge computing. The framework employs graph neural networks to capture spatial correlations among service regions and leverages Transformer-based self-attention to learn long-range temporal patterns from historical traffic observations. By decoupling spatial representation learning from temporal reasoning, the proposed approach provides an effective mechanism for large-scale spatiotemporal traffic modeling. Extensive experiments on a real-world cellular network dataset demonstrate that the proposed graph Transformer consistently outperforms recurrent graph-based baselines, including GCN-RNN, GCN-LSTM, and GCN-GRU models, across multiple forecasting horizons. The resulting forecasts enable more effective proactive resource provisioning and reduce overload risk compared with reactive management strategies. These results highlight the potential of graph-enhanced attention mechanisms for building intelligent and adaptive edge computing systems.
This framework introduces an adaptive graph learning module that dynamically infers meaningful connectivity relationships among traffic sensors—not relying on fixed geographic or distance-based assumptions—but instead leveraging real-time traffic correlations and node-level embeddings, enabling effective modeling of both localized spatial interactions and multi-scale temporal dependencies across varying prediction horizons.
Zhengxu Luan, Huan Wang, Miaobowen Wang et al.· Computers and artificial int...· 0 citations
In dynamic Vehicular Edge Computing (VEC) environments, rapidly changing vehicle mobility and traffic lead to fluctuating edge resource demands, challenging task offloading and scheduling. Accurate prediction of future traffic flow and traffic-state-derived workload representations is thus crucial for proactive resource management. To address the limitations of existing methods in short-term dynamic characterization, complex spatial interaction modeling, and heterogeneous target prediction, this paper proposes a Spatio-Temporal Graph Attention Network (TA-STGAT). The proposed model constructs multi-dimensional RSU-level state sequences from simulated trajectories generated on a real-world road network and separately forecasts vehicle flow within RSU coverage areas and the associated traffic-state-derived workload representation under a unified spatio-temporal modeling framework. By integrating gated dilated temporal convolutions with a topology-constrained multi-head graph attention mechanism, the model captures multi-scale temporal dependencies and nonlinear spatial correlations. Experimental results show that, compared with the best-performing baseline in terms of RMSE for each forecasting task, TA-STGAT reduces RMSE by 10.89% and 11.29% in workload-representation prediction and traffic flow prediction, respectively, demonstrating its effectiveness for short-term edge-state forecasting.
A novel method called adaptive diffused spatiotemporal graph convolution network (ADSTGCN) is proposed for accurate traffic flow prediction and achieves superior performance compared to other state-of-the-art methods.
Xiao Luo, Shanshan Wang, Shaobao Li et al.· Journal of Transportation En...· 0 citations
A Tensor-Evolving Graph with Temporal Separation Network (TEG-TSNet) for traffic flow forecasting is proposed, which constructs a unified spatial prior via graph Laplacian spectral embedding and introduces a time-conditioned structure generation paradigm.
Hongyan Wang, Linlong Chen· Journal of King Saud Univers...· 0 citations
TETRA is proposed, a hybrid spatio-temporal traffic forecasting model that integrates Graph Convolutional Networks (GCNs) with Extended Long Short-Term Memory (xLSTM) to capture complex multi-timescale temporal patterns, including congestion propagation and delayed recovery dynamics, which are not well represented by conventional recurrent models.
Norman Bereczki, Vilmos Simon· International Journal of Int...· 0 citations
Network-scale traffic speed prediction plays a central role in signal timing, routing, traffic control and congestion management. However, existing methods face challenges in capturing complex spatiotemporal dependencies among road segments, and in achieving computational efficiency in large-scale urban networks. To address these challenges, we propose an enhanced Dynamic Spatiotemporal Residual Network (DST-ResNet) framework for network-scale traffic speed prediction. First, the method employs a multi-scale grid partitioning strategy to segment urban road networks at varying levels of granularity, enabling precise predictions at both local and global scales. Second, a dynamic convolution mechanism is introduced to integrate real-time traffic data with static road network features, allowing the model to adaptively capture the dynamic and heterogeneous characteristics of urban traffic. Then, a spatiotemporal attention mechanism is incorporated to capture sequential dependencies in traffic patterns, improving the accuracy and robustness of long-term traffic speed prediction. Last, we conduct experiments based on GPS trajectory data from taxis in Xi’an city, China. The results demonstrate that DST-ResNet outperforms five baseline models. The model maintains stable performance across five spatial resolutions, and a mask-aware retraining strategy reduces the average MSE by 49.65% under missing observations. Ablation studies confirm the contributions of dynamic convolution and attention mechanisms, while residual units mitigate the vanishing gradient problem, facilitating deeper network training.
