Aug 2026· Frontiers of Computer Science· 0 citations· 17 references
TL;DR
Experimental results on the public PEMS04 and PEMS08 datasets demonstrate that the proposed ESDG-ALSTM model significantly improves forecasting accuracy, confirming that ESDG-ALSTM is more sensitive to abrupt events and multimodal evolution patterns and can effectively enhance prediction performance in complex traffic flow scenarios.
Abstract
To address the limitations of graph convolution-based traffic flow forecasting models, in which predefined static graphs are unable to adapt to the dynamic evolution of road networks and conventional LSTM models fail to finely characterize temporal dynamics, this study proposes an ESDG-ALSTM traffic flow forecasting model, namely an Embedded Sparse Dynamic Graph Convolutional LSTM model with Multi-Head Attention. First, a sparse dynamic graph convolution module is developed to integrate the physical topology and implicit semantic features of traffic flow data. Combined with a Top-K strategy, the adjacency matrix is dynamically reconstructed to enable adaptive capture of time-varying spatial dependencies while effectively suppressing long-tail noise. Next, an improved LSTM unit enhanced by multi-head attention is designed to model abrupt events and multimodal evolution patterns in traffic flow. Through multi-subspace projection and dynamic weight allocation, the hidden-state update process of the LSTM unit is restructured, thereby enabling deep extraction of traffic temporal features. Finally, a deeply coupled spatiotemporal feature network is constructed, in which graph convolution is further employed to refine the output states of the LSTM gates, achieving deep coupling and synchronized extraction of spatial and temporal features. Experimental results on the public PEMS04 and PEMS08 datasets demonstrate that the proposed model significantly improves forecasting accuracy, confirming that ESDG-ALSTM is more sensitive to abrupt events and multimodal evolution patterns and can effectively enhance prediction performance in complex traffic flow scenarios.
Traffic congestion has become a major obstacle to urban mobility and quality of life due to rapid urbanization and an increase in car ownership. Accurate traffic flow prediction is important for some intelligent transportation systems such as traffic guidance and signal control. The traditional methods including ARIMA and support vector machines cannot deal well with the complicated spatial and temporal correlations in large-scale traffic data. At present, most spatiotemporal graph neural networks (STGNNs) mainly depend on fixed graph structures which are not adaptive to the varying traffic conditions. We suggest a Dynamic Adaptive Attention Graph Learning based STGNN Transformer model (DASTGNN-Transformer) for multi-step traffic flow prediction. The model uses a multi-step Transformer decoder to identify global long-term temporal dependencies and obtains local spatiotemporal features by means of spatiotemporal graph convolutional encoders. Attention maps are generated from real-time data by a dynamic adaptive graph learning module to reflect the changing location relationships. Experiments conducted on PeMSD4 (containing 307 nodes) and PeMSD8 (with 170 nodes) datasets indicate that DASTGNN-Transformer outperforms the mainstream baselines such as STGCN, adaptive graph networks and iTransformer in terms of RMSE of 0.072–0.085, MAE of 0.061–0.079 and MAPE of 8.2%–9.5%. The ablation studies have confirmed the contributions of each module: the RMSE increases to 0.12–0.20 when the dynamic graph module is omitted. Furthermore, the robustness tests have demonstrated consistent performance even in the presence of noise (the RMSE only increases to 0.142 at a noise level of 0.5) and across different datasets. This research provides an integrated approach for dynamic spatiotemporal dependency modeling which significantly enhances the multi-step prediction accuracy.
Xiao-Li Wang· Engineering Research Express· 0 citations
Traffic flow prediction is a critical foundational problem in intelligent transportation systems. Although Large Language Model (LLM) has shown promising potential in time series modeling tasks in recent years, existing LLM-based methods generally overlook the inherent multi-scale characteristics of traffic flow data, which significantly limits their ability to capture complex spatio-temporal evolution patterns. To address this issue, this paper proposes a traffic flow forecasting framework named Multi-Scale Graph Convolution Enhanced Large Language Model (MSG-LLM). Firstly, the traffic flow series are decomposed based on frequency-domain analysis to identify periodic components, enabling the adaptive partitioning of the original series into multiple time scales. Subsequently, adaptive graph structures are constructed at different time scales, and graph convolution operations are introduced to fully characterize the correlation dependencies of traffic nodes during multi-scale spatio-temporal evolution. On this basis, a bidirectional multi-scale fusion module is designed to obtain comprehensive and consistent multi-scale representations through information fusion from fine-to-coarse and coarse-to-fine scales. Finally, the fused multi-scale spatio-temporal features are integrated into a partially frozen pre-trained large language model. By fine-tuning only task-specific parameters, this approach preserves the LLM’s general time series modeling capabilities while effectively reducing training costs and mitigating overfitting risks. Extensive experimental results on the PEMS04 and PEMS08 datasets demonstrate that the proposed method significantly outperforms existing mainstream models in both short-term and long-term traffic flow forecasting tasks, validating the effectiveness and strong generalization ability of MSG-LLM in modeling complex traffic 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
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
Traffic flow forecasting aims to predict future traffic states from historical observations of road networks. Although pretrained language models have demonstrated strong sequence modeling capacity, directly applying GPT-2 to traffic forecasting remains nontrivial because continuous spatiotemporal signals differ substantially from discrete text tokens. To address this issue, this paper proposes a node-tokenized GPT-2 framework and investigates attention adaptation and residual prediction mechanisms for traffic flow forecasting. Specifically, each traffic node is represented as a token by projecting its historical observations into the GPT-2 embedding space. Based on this representation, three attention mechanisms are compared, including causal attention, dense full attention, and dynamic Top-K sparse attention. In addition, a residual prediction strategy is introduced to model future traffic changes relative to the latest observed state. Experiments on the PeMS08 dataset show that GPT2-Causal-Res achieves the best performance, with an MAE of 15.3144, reducing the error by 3.47% compared with the direct prediction baseline. The results indicate that residual prediction improves forecasting accuracy, whereas simply modifying the attention structure does not necessarily lead to reliable modeling of traffic node interactions.
Ming-Zhu Gao, Ruo-Han Ning· International Conference on...· 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