Scholarlink

<p>Long-term Time Series Forecasting (LTSF) has garnered significant attention for its ability to predict extended time horizons. Commonly used deep decomposition models in time series forecasting excel at capturing complex patterns by decomposing the series into trend and seasonal components. However, these models often overlook the interactions between trends and seasons, potentially reducing prediction accuracy by failing to capture their correlations at multiple scales. To address this limitation, we propose the Multi-scale Trend-Seasonal Mixing Network (MTSMNet) for LTSF. MTSMNet introduces a multi-scale pattern mixture decomposition block to extract trend and seasonal patterns across various scales. These patterns are fed into a multi-scale trend-seasonal mixing layer to model the temporal correlations between trends and seasons. This mixing layer employs a unique Gated Recurrent Unit-based Encoder-Decoder structure at each scale, enhancing the accuracy of trend and seasonal patterns forecasts. Additionally, MTSMNet features a time-aware adaptive fusion module to dynamically integrate patterns across different scales over time. For the MSE value of prediction results on public datasets, MTSMNet has an average improvement of 29.32% in multivariable forecasting and 23.58% in univariable forecasting. The code is available at <a href="https://github.com/frfggv/TSF.git">https://github.com/frfggv/TSF.git</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

MTSMNet: a multi-scale trend-seasonal mixing network for long-term time series forecasting

  • 期刊论文

  • 发表日期：

    2025年8月21日

查看全文
  • Ruofan Feng,
  • Jiwei Qin,
  • Dezhi Sun,
  • Weilin Tang,
  • Xizhong Qin

摘要

Long-term Time Series Forecasting (LTSF) has garnered significant attention for its ability to predict extended time horizons. Commonly used deep decomposition models in time series forecasting excel at capturing complex patterns by decomposing the series into trend and seasonal components. However, these models often overlook the interactions between trends and seasons, potentially reducing prediction accuracy by failing to capture their correlations at multiple scales. To address this limitation, we propose the Multi-scale Trend-Seasonal Mixing Network (MTSMNet) for LTSF. MTSMNet introduces a multi-scale pattern mixture decomposition block to extract trend and seasonal patterns across various scales. These patterns are fed into a multi-scale trend-seasonal mixing layer to model the temporal correlations between trends and seasons. This mixing layer employs a unique Gated Recurrent Unit-based Encoder-Decoder structure at each scale, enhancing the accuracy of trend and seasonal patterns forecasts. Additionally, MTSMNet features a time-aware adaptive fusion module to dynamically integrate patterns across different scales over time. For the MSE value of prediction results on public datasets, MTSMNet has an average improvement of 29.32% in multivariable forecasting and 23.58% in univariable forecasting. The code is available at https://github.com/frfggv/TSF.git.