Scholarlink

<p>Long-term multivariate time series forecasting plays a critical role in applications ranging from electricity load management to financial risk assessment and climate monitoring. Current deep learning approaches face two fundamental limitations: existing Transformer variants struggle to effectively integrate multi-scale temporal patterns within unified architectures while balancing modeling capacity and computational cost, and time-frequency domain methods treat frequency decomposition and temporal modeling as separate stages, limiting unified multi-resolution learning potential. To address these limitations, we introduce WCDformer (Wavelet Cross-Scale Decoupled Transformer), a novel architecture that integrates wavelet-based decomposition with attention mechanisms through three key innovations: Wavelet Multi-Resolution Tokenization creates frequency-stratified representations via Haar decomposition, Temporal Pattern Decoupling Attention specializes query formation into trend, periodic, and residual branches for pattern-specific processing, and Inter-Scale Cross-Attention facilitates explicit information routing between temporal scales within a cohesive end-to-end framework. Extensive experiments across five benchmark datasets demonstrate WCDformer’s highly competitive performance against recent strong baselines, with a favorable accuracy-efficiency trade-off; multi-seed stability analysis on Weather further shows that several reported margins fall within the observed variability, particularly in challenging long-term forecasting scenarios. Our approach provides a framework for principled multi-scale temporal modeling, bridging classical signal processing with modern deep learning architectures for advanced forecasting systems. The source code is available at <a href="https://anonymous.4open.science/r/wcdformer-0207/">https://anonymous.4open.science/r/wcdformer-0207/</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

Beyond flat patching: wavelet tokens with cross-scale and decoupled attention for electricity load forecasting

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年8月1日

查看全文
  • Panfeng Ding

摘要

Long-term multivariate time series forecasting plays a critical role in applications ranging from electricity load management to financial risk assessment and climate monitoring. Current deep learning approaches face two fundamental limitations: existing Transformer variants struggle to effectively integrate multi-scale temporal patterns within unified architectures while balancing modeling capacity and computational cost, and time-frequency domain methods treat frequency decomposition and temporal modeling as separate stages, limiting unified multi-resolution learning potential. To address these limitations, we introduce WCDformer (Wavelet Cross-Scale Decoupled Transformer), a novel architecture that integrates wavelet-based decomposition with attention mechanisms through three key innovations: Wavelet Multi-Resolution Tokenization creates frequency-stratified representations via Haar decomposition, Temporal Pattern Decoupling Attention specializes query formation into trend, periodic, and residual branches for pattern-specific processing, and Inter-Scale Cross-Attention facilitates explicit information routing between temporal scales within a cohesive end-to-end framework. Extensive experiments across five benchmark datasets demonstrate WCDformer’s highly competitive performance against recent strong baselines, with a favorable accuracy-efficiency trade-off; multi-seed stability analysis on Weather further shows that several reported margins fall within the observed variability, particularly in challenging long-term forecasting scenarios. Our approach provides a framework for principled multi-scale temporal modeling, bridging classical signal processing with modern deep learning architectures for advanced forecasting systems. The source code is available at https://anonymous.4open.science/r/wcdformer-0207/.