全局-局部特征融合的多变量时序预测方法

doi:10.16451/j.cnki.issn1003-6059.202605002

摘要
图/表
参考文献
相关文章 (15)

全文: PDF (2600 KB) HTML (1 KB)
输出: BibTeX | EndNote (RIS)

摘要 Transformer在时间序列分析中展现巨大潜力,但其注意力机制常因聚合语义不相关的查询-键对而影响预测性能,同时,时间序列中存在的周期性、突发波动等复杂模式也增大建模难度.为此,文中提出全局-局部特征融合的多变量时序预测方法(Multivariate Time Series Forecasting Method with Global-Local Feature Fusion, MTS-GLFF).首先设计TopK-Transformer,根据可学习的传感器嵌入动态生成稀疏掩码,保留关键序列,进行特征聚合.在此基础上,构建双分支时间序列预测框架,包含全局分支网络和局部分支网络.全局分支网络通过跨变量注意力机制捕获全局的交互特征;局部分支网络采用多尺度架构,将时间序列分解成多粒度模式,精细化建模局部相关性.在10个基准数据集上的实验表明,MTS-GLFF在长短期预测任务中的性能均较优.

	服务

	把本文推荐给朋友
	加入我的书架
	加入引用管理器
	E-mail Alert
	RSS
	作者相关文章
	谢元
	强保华
	张世豪
	陈锐东
	陈礼睿
	张文辉

关键词 ：时间序列, 稀疏掩码, 注意力机制, 特征融合, Transformer

Abstract：The Transformer demonstrates considerable potential in time series analysis. However, the attention mechanism of the Transformer often aggregates semantically irrelevant query-key pairs, thereby resulting in the degradation of prediction performance. Moreover, complex patterns in time series, including periodicity and abrupt fluctuations, pose additional challenges for effective modeling. To address these issues, a multivariate time series forecasting method with global-local feature fusion(MTS-GLFF) is proposed. First, a TopK selection operator is designed. It dynamically generates sparse masks based on learnable sensor embeddings, thereby retaining key sequences for subsequent feature aggregation. Next, a dual-branch time series forecasting framework comprising global and local branch networks is constructed. The global branch captures global interaction features through a cross-variable attention mechanism, while the local branch adopts a multi-scale architecture that decomposes time series into multi-granularity patterns for fine-grained modeling of local dependencies. Experiments on 10 benchmark datasets demonstrate that MTS-GLFF achieves competitive performance.

Key words： Time Series Sparse Mask Attention Mechanism Feature Fusion Transformer

收稿日期: 2026-01-30

ZTFLH:

TP183

基金资助:国家自然科学基金项目(No.62262006)、广西重点研发计划项目(No.AB24010112,AB23026048)资助

通讯作者: 强保华,博士,教授,主要研究方向为大数据分析、时间序列预测、数据挖掘.E-mail:qiangbh@guet.edu.cn.

作者简介: 谢元,博士研究生,主要研究方向为时间序列表征学习、时间序列预测与异常检测、数据挖掘.E-mail:yuanxie@mails.guet.edu.cn.
张世豪,博士,讲师,主要研究方向为计算机视觉、时间序列分析.E-mail:zhangshihao@guet.edu.cn.
陈锐东,博士研究生,主要研究方向为深度学习、跨模态分析.E-mail:ruidongchen@mails.guet.edu.cn.
陈礼睿,博士研究生,主要研究方向为深度学习、计算机视觉.E-mail:LyricsChen@mails.guet.edu.cn.
张文辉,硕士,副教授,主要研究方向为数据挖掘、时间序列分析.E-mail:zhangwh@guet.edu.cn.

引用本文:

谢元, 强保华, 张世豪, 陈锐东, 陈礼睿, 张文辉. 全局-局部特征融合的多变量时序预测方法[J]. 模式识别与人工智能, 2026, 39(5): 397-414. XIE Yuan, QIANG Baohua, ZHANG Shihao, CHEN Ruidong, CHEN Lirui, ZHANG Wenhui. Multivariate Time Series Forecasting Method with Global-Local Feature Fusion. Pattern Recognition and Artificial Intelligence, 2026, 39(5): 397-414.

链接本文:

http://manu46.magtech.com.cn/Jweb_prai/CN/10.16451/j.cnki.issn1003-6059.202605002 或 http://manu46.magtech.com.cn/Jweb_prai/CN/Y2026/V39/I5/397

[1] 杨祎玥,伏潜,万定生.基于深度循环神经网络的时间序列预测模型.计算机技术与发展, 2017, 27(3): 35-38, 43.
(YANG Y Y, FU Q, WAN D S.A Prediction Model for Time Series Based on Deep Recurrent Neural Network. Computer Technology and Development, 2017, 27(3): 35-38, 43.)
[2] LI J L, LIU D S, MA G F, et al. DBFF-GRU: Dual-Branch Temporal Feature Fusion Network with Fast GRU for Multivariate Time Series Forecasting. Applied Intelligence, 2025, 55(7). DOI: 10.1007/s10489-025-06447-9.
[3] 杜永萍,牛晋宇,王陆霖,等.基于时间卷积注意力神经网络的序列推荐模型.模式识别与人工智能, 2022, 35(5): 472-480.
(DU Y P, NIU J Y, WANG L L, et al. Sequential Recommendation Model Based on Temporal Convolution Attention Neural Net-work. Pattern Recognition and Artificial Intelligence, 2022, 35(5): 472-480.)
[4] JO S H, WOO J, JEONG J H.TAMS: A CNN-Based Time Attention Network for Time Series Sensor Data with Feature Points of Bicycle Accident. Expert Systems with Applications, 2025, 272. DOI: 10.1016/j.eswa.2025.126739.
[5] ZENG A L, CHEN M X, ZHANG L, et al. Are Transformers Effective for Time Series Forecasting? Proceedings of the AAAI Confe-rence on Artificial Intelligence, 2023, 37(9): 11121-11128.
[6] LI Z, QI S Y, LI Y D, et al. Revisiting Long-Term Time Series Forecasting: An Investigation on Linear Mapping[C/OL].[2026-02-10]. https://arxiv.org/pdf/2305.10721.
[7] CHEN L, CHEN D H, SHANG Z J, et al. Multi-scale Adaptive Graph Neural Network for Multivariate Time Series Forecasting. IEEE Transactions on Knowledge and Data Engineering, 2023, 35(10): 10748-10761.
[8] KIM S, LEE T H, LEE J.TMF-GNN: Temporal Matrix Factorization-Based Graph Neural Network for Multivariate Time Series Forecasting with Missing Values. Expert Systems with Applications, 2025, 275. DOI: 10.1016/j.eswa.2025.127001.
[9] NIE Y Q, NGUYEN N H, SINTHONG P, et al. A Time Series Is Worth 64 Words: Long-Term Forecasting with Transformers[C/OL].[2026-02-10]. https://arxiv.org/pdf/2211.14730.14730.
[10] TANG W J, XIAO Y L, ZHOU Y Z, et al. PerioDformer: Periodic Disposition Enhanced Transformer for Times Series Forecasting. Knowledge-Based Systems, 2025, 327. DOI: 10.1016/j.knosys.2025.114088.
[11] SU G Y, GUAN Y P.MSDformer: An Autocorrelation Transformer with Multiscale Decomposition for Long-Term Multivariate Time Series Forecasting. Applied Intelligence, 2025, 55(2). DOI: 10.1007/s10489-024-06105-6.
[12] ZHANG P Y, YAN Y C, ZHANG X, et al. TransGNN: Harne-ssing the Collaborative Power of Transformers and Graph Neural Networks for Recommender Systems // Proc of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval. New York, USA: ACM, 2024: 1285-1295.
[13] BANDARA K, BERGMEIR C, HEWAMALAGE H.LSTM-MSNet: Leveraging Forecasts on Sets of Related Time Series with Multiple Seasonal Patterns. IEEE Transactions on Neural Networks and Learning Systems, 2021, 32(4): 1586-1599.
[14] LIU Q, WANG Q, WANG X Z.RVFL-LSTM: A Lightweight Model with Long-Short Term Memory for Time Series. Knowledge-Based Systems, 2025, 309. DOI: 10.1016/j.knosys.2024.112896.
[15] YU Z D, ZHANG C H, LIU J, et al. SKND-TSACNN: A Novel Time-Scale Adaptive CNN Framework for Fault Diagnosis of Rotating Machinery. Knowledge-Based Systems, 2023, 275. DOI: 10.1016/j.knosys.2023.110682.
[16] ZHANG Z Q, CHEN Y X, ZHANG D D, et al. CTFNet: Long-Sequence Time-Series Forecasting Based on Convolution and Time-Frequency Analysis. IEEE Transactions on Neural Networks and Learning Systems, 2024, 35(11): 16368-16382.
[17] GIAMPAOLO F, GATTA F, PREZIOSO E, et al. ENCODE-Ensemble Neural Combination for Optimal Dimensionality Encoding in Time-Series Forecasting. Information Fusion, 2023, 100. DOI: 10.1016/j.inffus.2023.101918.
[18] LIU M H, ZENG A L, CHEN M X, et al. SCINet: Time Series Modeling and Forecasting with Sample Convolution and Interaction // Proc of the 36th International Conference on Neural Information Processing Systems. Cambridge, USA: MIT Press, 2022: 5816-5828.
[19] DAS A, KONG W H, LEACH A, et al. Long-Term Forecasting with TiDE: Time-Series Dense Encoder[C/OL].[2026-02-10]. https://arxiv.org/pdf/2304.08424.
[20] JIANG H W, LIU D S, DING X Y, et al. TCM: An Efficient Lightweight MLP-Based Network with Affine Transformation for Long-Term Time Series Forecasting. Neurocomputing, 2025, 617. DOI: 10.1016/j.neucom.2024.128960.
[21] STITSYUK A, CHOI J. xPatch: Dual-Stream Time Series Forecasting with Exponential Seasonal-Trend Decomposition. Proceedings of the AAAI Conference on Artificial Intelligence, 2025, 39(19): 20601-20609.
[22] ZHOU H Y, ZHANG S H, PENG J Q, et al. Informer: Beyond Efficient Transformer for Long Sequence Time-Series Forecasting. Proceedings of the AAAI Conference on Artificial Intelligence, 2021, 35(12): 11106-11115.
[23] LIU S Z, YU H, LIAO C, et al. Pyraformer: Low-Complexity Pyramidal Attention for Long-Range Time Series Modeling and Forecasting[C/OL].[2026-02-10]. https://openreview.net/pdf?id=0EXmFzUn5I.
[24] ZHOU T, MA Z Q, WEN Q S, et al. FEDformer: Frequency Enhanced Decomposed Transformer for Long-Term Series Forecasting // Proc of the 39th International Conference on Machine Learning. San Diego, USA: JMLR, 2022: 27268-27286.
[25] WANG H X, MO Y P, XIANG K L, et al. CSformer: Combining Channel Independence and Mixing for Robust Multivariate Time Series Forecasting. Proceedings of the AAAI Conference on Artificial Intelligence, 2025, 39(20): 21090-21098.
[26] DING K Y, FAN F D, HOU C Y, et al. TimeMosaic: Temporal Heterogeneity Guided Time Series Forecasting via Adaptive Granularity Patch and Segment-Wise Decoding. Proceedings of the AAAI Conference on Artificial Intelligence, 2026, 40(25): 20790-20798.
[27] ZHANG Y H, YAN J C. Crossformer: Transformer Utilizing Cross-Dimension Dependency for Multivariate Time Series Forecasting[C/OL].[2026-02-10]. https://openreview.net/pdf?id=vSVLM2j9eie.
[28] LIU Y, HU T G, ZHANG H R, et al. iTransformer: Inverted Transformers Are Effective for Time Series Forecasting[C/OL].[2026-02-10]. https://arxiv.org/pdf/2310.06625.
[29] WU Z H, LIU Z J, LIN J, et al. Lite Transformer with Long-Short Range Attention[C/OL].[2026-02-10]. https://arxiv.org/pdf/2004.11886.
[30] ZHAO Y J, MA Z Q, ZHOU T, et al. GCformer: An Efficient Framework for Accurate and Scalable Long-Term Multivariate Time Series Forecasting // Proc of the 32nd ACM International Confe-rence on Information and Knowledge Management. New York, USA: ACM, 2023: 3464-3473.
[31] WANG H Q, PENG J, HUANG F H, et al. MICN: Multi-scale Local and Global Context Modeling for Long-Term Series Forecasting[C/OL].[2026-02-10]. https://openreview.net/pdf?id=zt53IDUR1U.
[32] PARK J, KANG H, HAN S, et al. Granularity Fusion Transfor-mer: Learning Multi-granularity Patterns for Time-Series Forecasting. Knowledge-Based Systems, 2025: 320. DOI: 10.1016/j.knosys.2025.113644.
[33] GUPTA A, DAR G, GOODMAN S, et al. Memory-Efficient Trans-formers via Top-k Attention // Proc of the 2nd Workshop on Simple and Efficient Natural Language Processing. Stroudsburg, USA: ACL, 2021: 39-52.
[34] WANG P C, WANG X, WANG F, et al. KVT: k-NN Attention for Boosting Vision Transformers // Proc of the 17th European Conference on Computer Vision. Berlin, Germany: Springer, 2022: 285-302.