Scholarlink

<p>Real-world multivariate time series often exhibit complex temporal dependencies across multiple time scales, making accurate modeling of both short-term fluctuations and long-term patterns a challenging task. Existing approaches face limitations: Transformer-based models effectively capture long-range dependencies but incur high computational costs, while convolution-based models are efficient yet struggle with long-range interactions. To address these limitations, we propose TCLformer, an encoder-decoder forecasting framework that integrates seasonal-trend decomposition, multi-scale dilated temporal convolution, and convolution-enhanced LogSparse self-attention (CELA). The decomposition strategy simplifies the modeling of non-stationary time series, while dilated temporal convolutions enable efficient multi-scale local pattern extraction. In addition, the CELA mechanism in the decoder selectively captures informative long-range dependencies with reduced redundancy, thereby&#xa0;supporting efficient long-horizon forecasting. Experimental results demonstrate that, for long-term forecasting, TCLformer achieves up to a 4.8% average MSE improvement over the strong baseline Robformer on the Electricity dataset, while for short-term forecasting, achieving up to a 14.5% average MSE improvement over FEDformer on the PEMS07 benchmark. Furthermore, empirical results show that TCLformer achieves a favorable balance between predictive accuracy and computational efficiency, requiring significantly fewer parameters and memory consumption&#xa0;compared to recent robustness-oriented Transformers.</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

TCLformer: enhancing multi-scale time series forecasting with temporal decomposition and convolution-enhanced LogSparse self-attention

  • 期刊论文

  • 发表日期：

    2026年3月23日

查看全文
  • Xiaohe Wu,
  • Kun Zhang,
  • Chenxi Cai,
  • Dianying Chen,
  • Yaodi Liu

摘要

Real-world multivariate time series often exhibit complex temporal dependencies across multiple time scales, making accurate modeling of both short-term fluctuations and long-term patterns a challenging task. Existing approaches face limitations: Transformer-based models effectively capture long-range dependencies but incur high computational costs, while convolution-based models are efficient yet struggle with long-range interactions. To address these limitations, we propose TCLformer, an encoder-decoder forecasting framework that integrates seasonal-trend decomposition, multi-scale dilated temporal convolution, and convolution-enhanced LogSparse self-attention (CELA). The decomposition strategy simplifies the modeling of non-stationary time series, while dilated temporal convolutions enable efficient multi-scale local pattern extraction. In addition, the CELA mechanism in the decoder selectively captures informative long-range dependencies with reduced redundancy, thereby supporting efficient long-horizon forecasting. Experimental results demonstrate that, for long-term forecasting, TCLformer achieves up to a 4.8% average MSE improvement over the strong baseline Robformer on the Electricity dataset, while for short-term forecasting, achieving up to a 14.5% average MSE improvement over FEDformer on the PEMS07 benchmark. Furthermore, empirical results show that TCLformer achieves a favorable balance between predictive accuracy and computational efficiency, requiring significantly fewer parameters and memory consumption compared to recent robustness-oriented Transformers.