| 103 | 0 | 100 |
| 下载次数 | 被引频次 | 阅读次数 |
【背景】时空交通流预测作为智能交通系统的重要基础能力，其在交通管理、调度优化与风险预警中具有关键作用。然而，在多数城市尤其是中小城市中，历史观测数据普遍稀缺，使得基于深度学习的时空预测模型难以训练；现有迁移学习方法由于城市间交通分布差异显著，往往存在泛化能力不足的问题。【目标】构建一种能够在目标场景小样本数据条件下保持稳定预测能力的跨场景迁移框架，并提升模型在不同场景的泛化性能。【方法】通过生成式预训练策略，以扩散超网络在参数空间中生成预测模型的完整参数，并结合参数序列化与区域提示以捕捉层间依赖与交通时空特征。【数据】使用四个大规模真实城市路网数据集进行评估，并采用仅包含目标城市三天数据的极少样本设定进行迁移测试。【结果】所构建框架在 15 min预测任务中取得RMSE 5.55和MAE 2.89的效果，整体性能显著优于所有对比基线。【结论】基于参数空间的生成式预训练能够有效缓解交通数据稀缺导致的建模困难，并显著提升跨城市迁移的鲁棒性。【应用】该框架可作为构建城市级交通基础模型的通用技术路径，为数据缺乏场景下的智慧交通管理与决策支持提供可行方案。
Abstract：[Background] Spatio-temporal traffic flow forecasting is fundamental to intelligent transportation systems， as it supports traffic management， operational optimization， and risk early warning. However， many cities， particularly small and medium-sized ones， suffer from severe data scarcity， which limits the effectiveness of data-driven methods. Moreover， existing transfer learning approaches often exhibit limited generalization capability due to substantial distribution shifts across cities. [Objective] To develop a cross-city migration framework that maintains stable forecasting performance under extremely limited target-city data and enhances the generalization ability across diverse urban environments. [Method] Generative pre-training strategy is adopted， where a diffusion-based hypernetwork generates the full parameter set of the forecasting model in parameter space， complemented by parameter serialization and region-aware prompts to capture inter-layer dependencies and city-specific characteristics. [Data] Four large-scale real-world urban road network datasets are used for evaluation， and migration tests are conducted under a few-shot setting with only three days of target-city data. [Result] The proposed framework achieves RMSE 5.55 and MAE 2.89 for 15-minute forecasting， significantly outperforming all baseline approaches. [Conclusion] Generative pre-training in parameter space effectively alleviates modeling difficulties caused by traffic data scarcity and notably enhances robustness in cross-city transfer. [Application] The framework serves as a generalizable pathway for building urban traffic foundation models and provides a practical solution for intelligent traffic management and decision support in data-limited scenarios.
[1] TANG Y， QU A， CHOW A H F， et al. Domain adversarial spatial-temporal network: a transferable framework for short-term traffic forecasting across cities[C]//Proceedings of the 31st ACM International Conference on Information & Knowledge Management. Atlanta: ACM， 2022: 1905-1915.
[2] HE T， BAO J， LI R， et al. What is the human mobility in a new city: transfer mobility knowledge across cities[C]//Proceedings of The Web Conference 2020. Taipei: ACM， 2020: 1355-1365.
[3] LU B， GAN X， ZHANG W， et al. Spatio-temporal graph few-shot learning with cross-city knowledge transfer[C]//Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. Washington DC: ACM， 2022: 1162-1172.
[4] LIU X， LIANG Y， HUANG C， et al. When do contrastive learning signals help spatio-temporal graph forecasting? [C]//Proceedings of the 30th International Conference on Advances in Geographic Information Systems. Seattle: ACM， 2022: 1-12.
[5] LI R， ZHONG T， JIANG X， et al. Mining spatio-temporal relations via self-paced graph contrastive learning[C]//Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. Washington DC: ACM， 2022: 936-944.
[6] DONG J X， LONG M S， WANG J M， et al. SimMTM: a simple pre-training framework for masked time-series modeling[C]//Advances in Neural Information Processing Systems 36. New Orleans: Neural Information Processing Systems Foundation， Inc. (NeurIPS)， 2023: 29996-30025.
[7] NIE Y Q， Nguyen N H， Sinthong P， et al. A time series is worth 64words: long-term forecasting with transformers [EB/OL]. (2022-11-27)[2025-11-03]. https://doi.org/10.48550/arXiv.2211.14730.
[8] SHAO Z， ZHANG Z， WANG F， et al. Pre-training enhanced spatial-temporal graph neural network for multivariate time series forecasting[C]//Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. Washington DC: ACM， 2022: 1567-1577.
[9] GAO H T， JIANG R H， DONG Z， et al. Spatial-temporal-decoupled masked pre-training for spatiotemporal forecasting [EB/OL]. (2023-12-01)[2025-11-03]. https://doi.org/10.48550/arXiv.2312.00516.
[10] YUAN Y， SHAO C Y， DING J T， et al. Spatio-temporal few-shot learning via diffusive neural network generation [EB/OL]. (2024-02-19)[2025-11-03]. https://doi.org/10.48550/arXiv.2402.11922.
[11] MA T， ZHAO Y， LI M， et al. TPLLM: a traffic prediction framework based on pretrained Large Language Models[J]. Applied Soft Computing， 2025， 184: 113840.
[12] WU H， ZHOU H， LONG M， et al. Interpretable weather forecasting for worldwide stations with a unified deep model[J]. Nature Machine Intelligence， 2023， 5(6): 602-611.
[13] FANG Z， LONG Q， SONG G， et al. Spatial-temporal graph ODE networks for traffic flow forecasting[C]//Proceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery & Data Mining. Virtual Event Singapore: ACM， 2021: 364-373.
[14] ZHENG C， FAN X， WANG C， et al. GMAN: a graph multi-attention network for traffic prediction[J]. Proceedings of the AAAI Conference on Artificial Intelligence， 2020， 34(1): 1234-1241.
[15] 朱子俊， 宋国华， 范鹏飞， 等. 基于趋势提取下多尺度时空残差网络的路段速度预测模型[J]. 交通运输工程与信息学报， 2023， 21(3): 74-85.
[16] XIAN J， YE Y， ZHANG W， et al. MDHGFN: Multiscale Dual Hypergraph Fusion Spatiotemporal Network for traffic flow prediction[J]. Chaos， Solitons & Fractals， 2025， 201: 117228.
[17] ZHANG H， LIN Z， XIE H， et al. Two-way heterogeneity model for dynamic spatiotemporal traffic flow prediction[J]. Knowledge-Based Systems， 2025， 320: 113635.
[18] LIU J， WANG Y， ZHU J， et al. A multilayer spatiotemporal correlation-aware graph attention network for traffic flow prediction[J/OL]. IEEE Transactions on Neural Networks and Learning Systems， 2025: 1-15. (2025-11-13)[2025-11-05] . DOI: 10.1109/TNNLS.2025.3630903.
[19] XU M X， DAI W R， LIU C M， et al. Spatial-temporal transformer networks for traffic flow forecasting[EB/OL]. (2020-06-09)[2025-11-05]. https://arxiv.org/abs/2001.02908.
[20] 姚鸿博， 张雪娜， 郑涵予， 等. 基于时空动态增强自注意力的交通预测[J]. 软件工程， 2026， 29(4): 1-7.
[21] LI Z， ZHOU J， LIN Z， et al. Dynamic spatial aware graph transformer for spatiotemporal traffic flow forecasting[J]. Knowledge-Based Systems， 2024， 297: 111946.
[22] 殷炽磊， 林之喆， 周腾， 等. 融合自适应图与时空Transformer的交通流预测模型[J]. 交通运输工程与信息学报， 2026， 24(1): 90-101.
[23] UNG H Q， NIU H， DAO M S， et al. xMTrans: temporal attentive cross-modality fusion transformer for long-term traffic prediction[C]//2024 25th IEEE International Conference on Mobile Data Management (MDM). Brussels: IEEE， 2024: 195-202.
[24] HO J， JAIN A， ABBEEL P. Denoising diffusion probabilistic models[EB/OL]. (2020-06-19)[2025-11-09]. https://arxiv.org/abs/2006.11239.
[25] ROMBACH R， BLATTMANN A， LORENZ D， et al. High-resolution image synthesis with latent diffusion models[C]//2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). New Orleans: IEEE， 2022: 10674-10685.
[26] FENG S， MIAO C， ZHANG Z， et al. Latent diffusion transformer for probabilistic time series forecasting[J]. Proceedings of the AAAI Conference on Artificial Intelligence， 2024， 38(11): 11979-11987.
[27] SU C， TIAN Y， SONG Y. Multimodal conditioned diffusive time series forecasting[EB/OL]. (2025-04-28)[2025-11-10]. https://arxiv.org/abs/2504.19669.
基本信息:
DOI：10.19961/j.cnki.1672-4747.2025.12.017
中图分类号:U495
引用信息:
[1]王梓赫,黄春翔,张彤彤,等.基于生成预训练增强的小样本交通流时空预测[J].交通运输工程与信息学报().DOI:10.19961/j.cnki.1672-4747.2025.12.017.
基金信息:
国家重点研发计划（2024YFB4303103）; 中国博士后科学基金面上项目（2025M774250）; 中央高校基本科研业务费
2026-04-30
2026-04-30
2026-04-30