Traffic flow prediction serves as a core component of intelligent transportation systems, rendering it indispensable for efficient traffic management and planning. To mitigate the limitations of existing methods in graph structure construction and attention computation processes, a traffic flow forecasting model that Integrating NAGC-FLA (node-adaptive graph convolution with focused linear attention) was proposed. The proposed model was primarily composed of three modules: a node-adaptive graph convolution module, a spatial focused linear attention module, and a temporal attention module. Initially, dynamic spatio-temporal features were extracted via embedding, and adaptive road network information was constructed to facilitate the learning of road network node features. Subsequently, the spatial focused linear attention module adopts linear attention to reduce computational complexity, while a focusing function was introduced to enhance feature extraction performance. Furthermore, the temporal attention module was utilized to capture non-linear temporal dependencies. Finally, node-adaptive graph convolution was employed to capture the unique variation patterns of each node, and gated fusion is performed with the output of the attention modules to dynamically adjust the weights assigned to the two types of outputs. Experimental validations are conducted on three real-world traffic datasets. In comparison with the optimal baseline model, the proposed model exhibits significant improvements in prediction accuracy: the three key evaluation metrics, namely mean absolute error, root mean square error and mean absolute percentage error, are reduced by an average of 2.49%, 1.42%, and 3.72%, respectively.
为此,现提出一种融合节点自适应图卷积和聚焦线性注意力模型来进行交通流量预测(traffic flow prediction model integrating node-adaptive graph convolution with focused linear attention,NAGC-FLA)。模型使用多维度嵌入输入来解构时空耦合特性,使其自适应学习交通模式特征。在此基础上,采用并行框架将时空注意力和图卷积进行门控融合。其中针对注意力提出空间聚焦线性注意力网络替换原有传统注意力网络,将二次计算复杂度降低为一次计算复杂度,并使用聚焦函数增强对于重点特征的提取能力,使用时间注意力提取时序依赖关系。针对图卷积采用节点自适应图卷积来动态学习路网空间特征。将NAGC-FLA在PEMS04、PEMS07、PEMS08数据集上与DDGCRN、PDFormer等11个模型进行比较,进而验证本文模型预测效果的优越性,以证明对于交通流量预测的有效性。
为了测试所提出的聚焦线性注意力是否真的能起到降低复杂度的作用,设计了2个变体模型进行实验。①no S FLA:不采用空间聚焦线性自注意力,而是在时间和空间维度都采用传统自注意力;②use ST FLA: 在采用空间聚焦线性自注意力的基础上,在时间维度也采用聚焦线性自注意力。
use S FLA即本文模型NAGC-FLA,使用空间聚焦线性自注意力和时间传统自注意力。表4为NAGC-FLA和变体模型在PEMS08数据集上的每轮训练时间和测试时间对比。为保证实验的严谨性,所有模型的batch_size和学习率均一致。由表4可知,相对于采用时空传统自注意力的no S FLA对比模型,所提出的空间聚焦线性注意力确实起到了降低计算复杂度的作用。而对于时空维度都采用聚焦线性自注意力的use ST FLA对比模型,发现时间维度上使用并不会起到降低计算复杂度的作用,反而会增大计算复杂度。这些证明了聚焦线性注意力在降低计算复杂度方面的有效性。
YanJiahe, LiHonghui, SunJing, et al. The development of spatio-temporal graph neural networks for traffic flow prediction: a review[J]. Computer Engineering and Applications, 2025, 61(21): 1-19.
[3]
KhanA, FoudaM M, DoD T, et al. Short-term traffic prediction using deep learning long short-term memory: taxonomy, applications, challenges, and future trends[J]. IEEE Access, 2023, 11: 94371-94391.
ZhongSonghao, LuYi, LiLi, et al. Long term traffic flow forecasting model based on cross period sparse technique[J]. Science Technology and Engineering, 2025, 25(24): 10468-10477.
[6]
WilliamsB M, HoelL A. Modeling and forecasting vehicular traffic flow as a seasonal ARIMA process: theoretical basis and empirical results[J]. Journal of Transportation Engineering, 2003, 129(6): 664-672.
JiangTing, YangLiu, LiuYalin, et al. Traffic flow prediction based on the dynamic spatial-temporal decomposition framework[J]. Science Technology and Engineering, 2025, 25(7): 3007-3017.
[9]
PascaleA, NicoliM. Adaptive Bayesian network for traffic flow prediction[C]// IEEE Statistical Signal Processing Workshop (SSP). Nice: IEEE, 2011: 177-180.
[10]
SunB, ChengW, GoswamiP, et al. Flow-aware WPT k-nearest neighbours regression for short-term traffic prediction[C]// IEEE Symposium on Computers and Communications (ISCC). Heraklion: IEEE, 2017: 48-53.
[11]
GengZ, XuJ, WuR, et al. STGAFormer: spatial-temporal gated attention transformer based graph neural network for traffic flow forecasting[J]. Information Fusion, 2024, 105: 102228.
[12]
LiuQ, LiuT, CaiY, et al. Explanatory prediction of traffic congestion propagation mode: a self-attention based approach[J]. Physica A: Statistical Mechanics Its Applications, 2021, 573: 125940.
[13]
MaX, DaiZ, HeZ, et al. Learning traffic as images: a deep convolutional neural network for large-scale transportation network speed prediction[J]. Sensors, 2017, 17(4): 818.
[14]
JinG, LiangY, FangY, et al. Spatio-temporal graph neural networks for predictive learning in urban computing: a survey[J]. IEEE Transactions on Knowledge Data Engineering, 2023, 36(10): 5388-408.
CuiJianxun, YaoJia, ZhaoBoyuan. Review on short-term traffic flow prediction methods based on deep learning[J]. Journal of Traffic and Transportation Engineering, 2024, 24(2): 50-64.
[17]
DaiG, KongW, LiuY, et al. Multi-perspective convolutional neural networks for citywide crowd flow prediction[J]. Applied Intelligence, 2023, 53(8): 8994-9008.
[18]
YuanQ, WangJ, HanY, et al. DAGCAN: decoupled adaptive graph convolution attention network for traffic forecasting[J]. IEEE Transactions on Intelligent Transportation Systems, 2025, 26(3): 3513-3526.
[19]
YuB, YinH, ZhuZ J. Spatio-temporal graph convolutional networks: a deep learning framework for traffic forecasting[J]. arXiv Preprint arXiv: 04875, 2017.
GuoS, LinY, FengN, et al. Attention based spatial-temporal graph convolutional networks for traffic flow forecasting[C]// Proceedings of the AAAI Conference on Artificial Intelligence. Honolulu: AAAI, 2019: 922-929.
[22]
WuZ, PanS, LongG, et al. Graph WaveNet for deep spatial-temporal graph modeling[C]// 28th International Joint Conference on Artificial Intelligence. Macao: IJCAI, 2019: 1907-1913.
[23]
BaiL, YaoL, LiC, et al. Adaptive graph convolutional recurrent network for traffic forecasting[J]. Advances in Neural Information Processing Systems, 2020, 33: 17804-17815.
[24]
LiF, FengJ, YanH, et al. Dynamic graph convolutional recurrent network for traffic prediction: benchmark and solution[J]. ACM Transactions on Knowledge Discovery from Data, 2023, 17(1): 1-21.
[25]
JiangW, LuoJ, HeM, et al. Graph neural network for traffic forecasting: the research progress[J]. ISPRS International Journal of Geo-Information, 2023, 12(3): 100.
[26]
LiuJ, KangY, LiH, et al. STGHTN: spatial-temporal gated hybrid transformer network for traffic flow forecasting[J]. Applied Intelligence, 2023, 53(10): 12472-12488.
[27]
XingZ, HuangM, LiW, et al. Spatial linear transformer and temporal convolution network for traffic flow prediction[J]. Scientific Reports, 2024, 14(1): 4040.
[28]
AngelosK, ApoorvV, NikolaosP, et al. Transformers are RNNs: fast autoregressive transformers with linear attention[C]// International Conference on Machine Learning (ICML). Online: IMLS, 2020: 5156-5165.
[29]
ShaoZ, ZhangZ, WangF, et al. Spatial-temporal identity: a simple yet effective baseline for multivariate time series forecasting[C]// Proceedings of the 31st ACM International Conference on Information & Knowledge Management. Atlanta: ACM, 2022: 4454-4458.
[30]
JiangJ, HanC, ZhaoW X, et al. Pdformer: propagation delay-aware dynamic long-range transformer for traffic flow prediction[C]// Proceedings of the AAAI Conference on Artificial Intelligence. Washington, D.C.: AAAI, 2023: 4365-4373.
[31]
WengW, FanJ, WuH, et al. A decomposition dynamic graph convolutional recurrent network for traffic forecasting[J]. Pattern Recognition, 2023, 142: 109670.
HeWenwu, PeiBoyu, LiYating, et al. Traffic flow forecasting based on bi-directional adaptive gating graph convolutional networks[J]. Journal of Transportation Systems Engineering and Information Technology, 2023, 23(1): 187-197.
[34]
KongW, GuoZ, LiuY. Spatio-temporal pivotal graph neural networks for traffic flow forecasting[C]// Proceedings of the AAAI Conference on Artificial Intelligence. Vancouver: AAAI, 2024: 8627-8635.
[35]
NingX J. Linear attention is enough in spatial-temporal forecasting[J]. arXiv Preprint arXiv: 09158, 2024.