Scholarlink

<p>Continuous Wavelet Transform (CWT) offers fine-grained time–frequency analysis but incurs substantial computational overhead, whereas Discrete Wavelet Transform (DWT), though efficient, struggles to represent features across continuous scales. To address this limitation, we propose LTF-Net, a lightweight time–frequency fusion framework. LTF-Net introduces a learnable wavelet kernel (LCWT) to construct a pseudo-continuous wavelet decomposition, effectively preserving rich time–frequency structures while markedly reducing computational cost. Complemented by global frequency representations extracted via the Fast Fourier Transform (FFT) and a cross-domain attention mechanism that integrates the wavelet and Fourier branches, the model jointly captures long-term periodic patterns and local non-stationary dynamics. Extensive experiments on multiple long-horizon forecasting benchmarks demonstrate the superiority of LTF-Net. Across ECL, Weather, ETTh1, ETTh2, ETTm1, and ETTm2, LTF-Net consistently outperforms state-of-the-art approaches, achieving nearly 40% MSE improvement over TimesNet on ETTm2. Competitive performance is also observed on the Traffic dataset. Overall, LTF-Net exhibits strong multi-scale modeling capability and robust generalization in long-term time series forecasting.</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

Lightweight time–frequency representation learning for time series forecasting

  • 期刊论文

  • 发表日期：

    2026年3月30日

查看全文
  • Jian Xiao,
  • Dongrui Han,
  • Xin Hu

摘要

Continuous Wavelet Transform (CWT) offers fine-grained time–frequency analysis but incurs substantial computational overhead, whereas Discrete Wavelet Transform (DWT), though efficient, struggles to represent features across continuous scales. To address this limitation, we propose LTF-Net, a lightweight time–frequency fusion framework. LTF-Net introduces a learnable wavelet kernel (LCWT) to construct a pseudo-continuous wavelet decomposition, effectively preserving rich time–frequency structures while markedly reducing computational cost. Complemented by global frequency representations extracted via the Fast Fourier Transform (FFT) and a cross-domain attention mechanism that integrates the wavelet and Fourier branches, the model jointly captures long-term periodic patterns and local non-stationary dynamics. Extensive experiments on multiple long-horizon forecasting benchmarks demonstrate the superiority of LTF-Net. Across ECL, Weather, ETTh1, ETTh2, ETTm1, and ETTm2, LTF-Net consistently outperforms state-of-the-art approaches, achieving nearly 40% MSE improvement over TimesNet on ETTm2. Competitive performance is also observed on the Traffic dataset. Overall, LTF-Net exhibits strong multi-scale modeling capability and robust generalization in long-term time series forecasting.