Scholarlink

<p>Long-term time series forecasting is a critical task in computational intelligence, playing a vital role in real-time decision-making systems. However, modeling large-scale multivariate time series requires extreme computational and memory efficiency. While Transformer-based models capture long-range dependencies, their quadratic complexity creates a severe memory bottleneck, significantly limiting real-time throughput. Conversely, highly efficient linear models fail to represent intricate nonlinear local dynamics. Moreover, existing multi-scale frameworks commonly employ average pooling, which acts as an imperfect smoothing filter and introduces structural distortion—leading to the irreversible loss of fine-grained temporal variations. To overcome these challenges, we propose DPWMixer, a computationally efficient dual-path architecture. The framework is built upon an Efficient Haar Wavelet Pyramid that replaces traditional pooling, utilizing orthogonal decomposition to explicitly disentangle macro-trends and local fluctuations while preserving structural integrity. To process these components, we design a Dual-Path Trend Mixer that integrates a global linear mapping for macro-trend anchoring and a flexible patch-based MLP-Mixer for micro-dynamic evolution. Finally, an adaptive multi-scale fusion module integrates predictions from diverse temporal resolutions, weighted by channel characteristics to optimize synthesis. Extensive experiments on eight public benchmarks demonstrate that our method achieves highly competitive forecasting accuracy against state-of-the-art baselines. By scaling linearly with a minimal computational constant, DPWMixer maintains an ultra-low memory footprint and fast training speed, making it exceptionally suited for large-scale data processing. The code is available at <a href="https://github.com/hit636/DPWMixer">https://github.com/hit636/DPWMixer</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

Dpwmixer: dual-path wavelet mixer for long-term time series forecasting

  • 期刊论文

  • 发表日期：

    2026年5月19日

查看全文
  • Qianyang Li,
  • Xingjun Zhang,
  • Shaoxun Wang,
  • Jia Wei

摘要

Long-term time series forecasting is a critical task in computational intelligence, playing a vital role in real-time decision-making systems. However, modeling large-scale multivariate time series requires extreme computational and memory efficiency. While Transformer-based models capture long-range dependencies, their quadratic complexity creates a severe memory bottleneck, significantly limiting real-time throughput. Conversely, highly efficient linear models fail to represent intricate nonlinear local dynamics. Moreover, existing multi-scale frameworks commonly employ average pooling, which acts as an imperfect smoothing filter and introduces structural distortion—leading to the irreversible loss of fine-grained temporal variations. To overcome these challenges, we propose DPWMixer, a computationally efficient dual-path architecture. The framework is built upon an Efficient Haar Wavelet Pyramid that replaces traditional pooling, utilizing orthogonal decomposition to explicitly disentangle macro-trends and local fluctuations while preserving structural integrity. To process these components, we design a Dual-Path Trend Mixer that integrates a global linear mapping for macro-trend anchoring and a flexible patch-based MLP-Mixer for micro-dynamic evolution. Finally, an adaptive multi-scale fusion module integrates predictions from diverse temporal resolutions, weighted by channel characteristics to optimize synthesis. Extensive experiments on eight public benchmarks demonstrate that our method achieves highly competitive forecasting accuracy against state-of-the-art baselines. By scaling linearly with a minimal computational constant, DPWMixer maintains an ultra-low memory footprint and fast training speed, making it exceptionally suited for large-scale data processing. The code is available at https://github.com/hit636/DPWMixer.