Scholarlink

<p>Transformer-based models have exhibited superior performance in the field of multivariate long-sequence time series forecasting. The Informer model, which adopts the probabilistic sparse self-attention mechanism, presents advantages in computational complexity compared to other models. However, the correlations between variables and multi-scale local features are difficult to be effectively captured by Informer, which results in suboptimal prediction accuracy. To improve prediction accuracy and further reduce computational complexity, a novel Informer-based model, named TACDformer, is proposed in this paper. Firstly, a multi-scale feature distillation layer is designed. This layer can be utilized to extract multi-scale local features that are difficult to be captured by the self-attention mechanism, and to perform adaptive downsampling while fully capturing temporal dependencies and inter-sequence dependencies. Furthermore, an encoder-only architecture is adopted by the model, whereby the negative impacts of the stacked decoder architecture and the historical time series used as decoder inputs on the prediction results are avoided. Since the features output by the encoder cannot be fully learned through simple linear mapping, an output layer based on the temporal attention mechanism and channel attention mechanism is designed. Model performance is enhanced and computational complexity is reduced simultaneously by this layer. Finally, the reversible instance normalization method is introduced on the basis of the Informer model; by performing normalization and denormalization during the data input and output phases respectively, the prediction bias caused by data distribution differences is reduced. To validate the effectiveness of TACDformer, extensive experiments were conducted on seven benchmark datasets spanning four domains: electricity, finance, energy, and weather. The results demonstrate that TACDformer outperforms eight baseline models, including TimeMixer and iTransformer, in terms of MAE and MSE metrics. Meanwhile, compared with Informer, the computational complexity and memory usage of TACDformer are reduced by 33.2% and 63.4%, respectively, and its running time is shortened by 77.2%. Code is available at: <a href="https://github.com/hclxk-hzy/TACDformer">https://github.com/hclxk-hzy/TACDformer</a>.</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

TACDformer: an improved informer-based model for accurate multivariate long-term time series forecasting

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年4月9日

查看全文
  • Zeyu Hu,
  • Yuan Jia,
  • Wu Le,
  • Zhenhong Jia,
  • Congbing He,
  • Huihui Fan,
  • Jie Meng

摘要

Transformer-based models have exhibited superior performance in the field of multivariate long-sequence time series forecasting. The Informer model, which adopts the probabilistic sparse self-attention mechanism, presents advantages in computational complexity compared to other models. However, the correlations between variables and multi-scale local features are difficult to be effectively captured by Informer, which results in suboptimal prediction accuracy. To improve prediction accuracy and further reduce computational complexity, a novel Informer-based model, named TACDformer, is proposed in this paper. Firstly, a multi-scale feature distillation layer is designed. This layer can be utilized to extract multi-scale local features that are difficult to be captured by the self-attention mechanism, and to perform adaptive downsampling while fully capturing temporal dependencies and inter-sequence dependencies. Furthermore, an encoder-only architecture is adopted by the model, whereby the negative impacts of the stacked decoder architecture and the historical time series used as decoder inputs on the prediction results are avoided. Since the features output by the encoder cannot be fully learned through simple linear mapping, an output layer based on the temporal attention mechanism and channel attention mechanism is designed. Model performance is enhanced and computational complexity is reduced simultaneously by this layer. Finally, the reversible instance normalization method is introduced on the basis of the Informer model; by performing normalization and denormalization during the data input and output phases respectively, the prediction bias caused by data distribution differences is reduced. To validate the effectiveness of TACDformer, extensive experiments were conducted on seven benchmark datasets spanning four domains: electricity, finance, energy, and weather. The results demonstrate that TACDformer outperforms eight baseline models, including TimeMixer and iTransformer, in terms of MAE and MSE metrics. Meanwhile, compared with Informer, the computational complexity and memory usage of TACDformer are reduced by 33.2% and 63.4%, respectively, and its running time is shortened by 77.2%. Code is available at: https://github.com/hclxk-hzy/TACDformer.