Scholarlink

<p>Long-term time series forecasting remains a challenging task due to the complex entanglement of temporal patterns and the inherent limitations of traditional attention mechanisms in capturing long-range dependencies efficiently. To address these challenges, we propose a novel framework named Decompose-Disentangle-Decode Time Series Forecasting (D<sup>3</sup>TSF), which aims to establish a lightweight yet high-precision forecasting paradigm. First, to mitigate the adverse effects of non-stationary data distribution shifts, we employ Reversible Instance Normalization (RevIN). Subsequently, we introduce a dynamically weighted decomposition mechanism based on Exponential Moving Average (EMA) to overcome the lag inherent in traditional Simple Moving Average (SMA) methods, effectively decoupling the time series into seasonal and trend components. Following a “divide-and-conquer” strategy, we design a dual-stream architecture: the innovative Patch-wise State Space Model (PwSSM) is utilized to capture fine-grained local features within highly fluctuating seasonal series, while a uniquely designed Sparse Linear MLP (SLMLP) extracts dominant long-term linear features from trend series. Extensive experiments across eight benchmark datasets demonstrate the superiority of D<sup>3</sup>TSF. Out of 64 evaluation metrics, our model achieves 18 best and 22&#xa0;s-best results, comprehensively outperforming state-of-the-art (SOTA) baselines such as EIAformer. Notably, D<sup>3</sup>TSF yields a significant performance improvement on the Weather dataset and reduces the MSE by 6.2% on ETTm1 (720 steps) with iTransformer, while exhibiting exceptional stability against noise (SNR 30&#xa0;dB) and error accumulation in long-horizon forecasting. These results validate that D<sup>3</sup>TSF offers a robust and computationally efficient solution for complex real-world forecasting scenarios.</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

Decompose-Disentangle-Decode Time Series Forecasting (D3TSF): dual-scale time series forecasting via seasonality of patch-wise state space and sparse trend learning

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月7日

查看全文
  • Shenglin Liu

摘要

Long-term time series forecasting remains a challenging task due to the complex entanglement of temporal patterns and the inherent limitations of traditional attention mechanisms in capturing long-range dependencies efficiently. To address these challenges, we propose a novel framework named Decompose-Disentangle-Decode Time Series Forecasting (D3TSF), which aims to establish a lightweight yet high-precision forecasting paradigm. First, to mitigate the adverse effects of non-stationary data distribution shifts, we employ Reversible Instance Normalization (RevIN). Subsequently, we introduce a dynamically weighted decomposition mechanism based on Exponential Moving Average (EMA) to overcome the lag inherent in traditional Simple Moving Average (SMA) methods, effectively decoupling the time series into seasonal and trend components. Following a “divide-and-conquer” strategy, we design a dual-stream architecture: the innovative Patch-wise State Space Model (PwSSM) is utilized to capture fine-grained local features within highly fluctuating seasonal series, while a uniquely designed Sparse Linear MLP (SLMLP) extracts dominant long-term linear features from trend series. Extensive experiments across eight benchmark datasets demonstrate the superiority of D3TSF. Out of 64 evaluation metrics, our model achieves 18 best and 22 s-best results, comprehensively outperforming state-of-the-art (SOTA) baselines such as EIAformer. Notably, D3TSF yields a significant performance improvement on the Weather dataset and reduces the MSE by 6.2% on ETTm1 (720 steps) with iTransformer, while exhibiting exceptional stability against noise (SNR 30 dB) and error accumulation in long-horizon forecasting. These results validate that D3TSF offers a robust and computationally efficient solution for complex real-world forecasting scenarios.