Scholarlink

<p>Transformer-based models have demonstrated remarkable advancements in Long-term Time Series Forecasting (LTSF) through self-attention mechanisms to capture long-range temporal dependencies. However, the self-attention mechanisms emphasize intra-sequence temporal correlations while neglecting inter-sequence relationships and the importance of different channels. Besides, the Transformer’s fixed additive fusion of input features and Feed-Forward Network (FFN) outputs lacks an adaptive weighting mechanism, which inherently limits its representational capacity. To address these limitations, we propose a Transformer-based model with hybrid attention and Kolmogorov-Arnold Networks (KANs), named as HAKformer. The HAKformer framework integrates three components: 1) A Hybrid Temporal-Channel Attention (HTCA) mechanism that simultaneously captures temporal external correlations and channels importance through dual attention pathways; 2) A Self-learning Fusion (SLF) mechanism that adaptively the optimal ratio between input and FFN using learnable weighting paprameters; 3) A KAN that enhances nonlinear pattern recongnition through spline-based activation functions. Experimental results on nine real-world time series datasets indicate that HAKformer achieves state-of-the-art performance in long-term forecasting. The code is made available at: <a href="https://github.com/Lvjiaqi009/HAKformer.">https://github.com/Lvjiaqi009/HAKformer.</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

HAKformer: an efficient transformer with hybrid temporal-channel attention and Kolmogorov-Arnold networks for time series forecasting

  • 期刊论文

  • 发表日期：

    2026年4月13日

查看全文
  • Jiaqi Lv,
  • Zhiyong An,
  • Yongliang Wang

摘要

Transformer-based models have demonstrated remarkable advancements in Long-term Time Series Forecasting (LTSF) through self-attention mechanisms to capture long-range temporal dependencies. However, the self-attention mechanisms emphasize intra-sequence temporal correlations while neglecting inter-sequence relationships and the importance of different channels. Besides, the Transformer’s fixed additive fusion of input features and Feed-Forward Network (FFN) outputs lacks an adaptive weighting mechanism, which inherently limits its representational capacity. To address these limitations, we propose a Transformer-based model with hybrid attention and Kolmogorov-Arnold Networks (KANs), named as HAKformer. The HAKformer framework integrates three components: 1) A Hybrid Temporal-Channel Attention (HTCA) mechanism that simultaneously captures temporal external correlations and channels importance through dual attention pathways; 2) A Self-learning Fusion (SLF) mechanism that adaptively the optimal ratio between input and FFN using learnable weighting paprameters; 3) A KAN that enhances nonlinear pattern recongnition through spline-based activation functions. Experimental results on nine real-world time series datasets indicate that HAKformer achieves state-of-the-art performance in long-term forecasting. The code is made available at: https://github.com/Lvjiaqi009/HAKformer.