Scholarlink

<p>In recent years, the proliferation of large-scale Internet of Things (IoT) devices has resulted in the generation of vast amounts of time-series data, thereby increasing the focus on multivariate time series prediction research. To effectively process IoT data, it is crucial to develop more advanced multivariate time series prediction models. Currently, the most sophisticated models predominantly employ Channel Independence (CI) strategies. However, CI strategies demonstrate limited generalization capabilities for unseen instances and fail to account for fundamental interactions between channels. Conversely, the Channel Dependency (CD) strategy indiscriminately merges all channels, which can lead to over-smoothing issues and consequently restrict prediction accuracy. To overcome these challenges, we propose an innovative adaptive differential channel method that synergizes the data expansion benefits of the CI strategy with the anti-correlation forgetting capabilities of the CD strategy. Utilizing this method, we introduce ADCformer, which adeptly captures inter-channel correlation information through the adaptive differential channel method and models long-term features via an attention mechanism. Experimental results indicate that our method enhances long-term time series prediction performance by 4.5% compared to state-of-the-art models.</p>

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

ADCformer: multivariate time series forecasting with adaptive differential channels transformer

  • 期刊论文

  • 发表日期：

    2026年4月7日

查看全文
  • Yutao Xia,
  • Yue Zhou,
  • Xiaofang Hu,
  • Shukai Duan

摘要

In recent years, the proliferation of large-scale Internet of Things (IoT) devices has resulted in the generation of vast amounts of time-series data, thereby increasing the focus on multivariate time series prediction research. To effectively process IoT data, it is crucial to develop more advanced multivariate time series prediction models. Currently, the most sophisticated models predominantly employ Channel Independence (CI) strategies. However, CI strategies demonstrate limited generalization capabilities for unseen instances and fail to account for fundamental interactions between channels. Conversely, the Channel Dependency (CD) strategy indiscriminately merges all channels, which can lead to over-smoothing issues and consequently restrict prediction accuracy. To overcome these challenges, we propose an innovative adaptive differential channel method that synergizes the data expansion benefits of the CI strategy with the anti-correlation forgetting capabilities of the CD strategy. Utilizing this method, we introduce ADCformer, which adeptly captures inter-channel correlation information through the adaptive differential channel method and models long-term features via an attention mechanism. Experimental results indicate that our method enhances long-term time series prediction performance by 4.5% compared to state-of-the-art models.