Scholarlink

This study focuses on generating long-term time series data. Existing deep learning-based methods for generating long-term sequences face two main challenges: limited ability in capturing long-term temporal relationships and gradient problems during training. To address these issues, we propose a long-term time series generative adversarial network (LTTS-GAN) by exploiting a multi-channel progressive decomposition generator. Both the generator and discriminator in LTTS-GAN are built upon the transformer encoder structure. In addition, we enhance the generator’s architecture by incorporating multiple channels of trend information from real data, aiming to improve the quality of the generated data. Furthermore, we employ the Auto-Correlation mechanism to identify period-wise relationships while introducing a global-local attention mechanism to balance local and long-range information. As a result, LTTS-GAN demonstrates superior capability in capturing long-term sequence features compared to other models. We validate the effectiveness of LTTS-GAN through a real-world classification application, confirming its performance across various scenarios.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

LTTS-GAN: A Long-Term Time Series Generative Adversarial Network

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Xiangyu Cui,
  • Xianping Ma,
  • Man-On Pun,
  • Zhimin Cheng

摘要

This study focuses on generating long-term time series data. Existing deep learning-based methods for generating long-term sequences face two main challenges: limited ability in capturing long-term temporal relationships and gradient problems during training. To address these issues, we propose a long-term time series generative adversarial network (LTTS-GAN) by exploiting a multi-channel progressive decomposition generator. Both the generator and discriminator in LTTS-GAN are built upon the transformer encoder structure. In addition, we enhance the generator’s architecture by incorporating multiple channels of trend information from real data, aiming to improve the quality of the generated data. Furthermore, we employ the Auto-Correlation mechanism to identify period-wise relationships while introducing a global-local attention mechanism to balance local and long-range information. As a result, LTTS-GAN demonstrates superior capability in capturing long-term sequence features compared to other models. We validate the effectiveness of LTTS-GAN through a real-world classification application, confirming its performance across various scenarios.