Please submit manuscripts in either of the following two submission systems

    ScholarOne Manuscripts

  • ScholarOne
  • 勤云稿件系统

  • 登录

Search by Issue

  • 2026 Vol.33
  • 2025 Vol.32
  • 2024 Vol.31
  • 2023 Vol.30
  • 2022 Vol.29
  • 2021 Vol.28
  • 2020 Vol.27
  • 2019 Vol.26
  • 2018 Vol.25
  • 2017 Vol.24
  • 2016 vol.23
  • 2015 vol.22
  • 2014 vol.21
  • 2013 vol.20
  • 2012 vol.19
  • 2011 vol.18
  • 2010 vol.17
  • 2009 vol.16
  • No.1
  • No.2

Supervised by Ministry of Industry and Information Technology of The People's Republic of China Sponsored by Harbin Institute of Technology Editor-in-chief Yu Zhou ISSNISSN 1005-9113 CNCN 23-1378/T

期刊网站二维码
微信公众号二维码
Related citation:
【HTML】   【PDF download】   【View/Add Comment】
Back Issue    Advanced Search
This paper has been browsed  65 times  ,  downloaded  77 times  
Shared by: Wechat More
A time series imputation algorithm based on attention-cue Wasserstein GAN with spectral normalization imputation
Author NameAffiliationPostcode
Yiyang Nan School of Electrical Engineering, Dalian Jiaotong University, Dalian 116028, China 116028
Wei Ba* School of Electrical Engineering, Dalian Jiaotong University, Dalian 116028, China 116028
Zhuang Li School of Electrical Engineering, Dalian Jiaotong University, Dalian 116028, China 116028
Mingyang Luo School of Electrical Engineering, Dalian Jiaotong University, Dalian 116028, China 116028
Abstract:
When the time series data contains continuous missing fragments, it is much more difficult to fill in the data than in the case of sporadic missing points. In the long deletion interval, it is often difficult for generators in traditional generative adversarial networks (GANs) to obtain sufficient distribution information from incomplete sequences. The discriminator is also affected because some inputs are missing or replaced, which may introduce deviations in the evaluation process. These problems will weaken the stability of training and the accuracy of filling. To address this problem, this study proposes an attention prompt with spectrum normalisation Wasserstein GAN (ACWGAN-SN) for time series data filling. The method introduces a dynamic prompt matrix to help the generator leverage context information around the missing interval. The matrix is generated by the multi-head attention mechanism, which can capture the long-term time dependence in the sequence observation; then, through the mask matrix, the generated context prompt information is placed only in the missing position. This method allows the generator to reconstruct continuous missing fragments with the support of more time context information, instead of relying mainly on noise. In terms of discriminators, the method applies spectral normalisation technology to the weight matrices of each layer. By controlling the maximum singular value of the weight matrix, the discriminator can better meet the 1-Lipschitz constraint in the Wasserstein framework, making the anti-training process more stable. Experimental results on two benchmark datasets show that at a 30%-50% deletion rate, the error of ACWGAN-SN is lower than that of the five reference models, and its RMSE, MASE and SMAPE indicators are reduced by 17.8%, 20.3%, and 18.8%, respectively. The results of the ablation experiment also confirm that the dynamic prompt matrix and spectral normalisation play an important role in improving the accuracy of filling and training stability.
Key words:  Time Series Imputation  Generative Adversarial Networks  Attention Mechanism  Spectral Normalization
DOI:10.11916/j.issn.1005-9113.2026054
Clc Number:TP274

LINKS