Scholarlink

<p>Long-horizon Time Series Forecasting (LTSF) is essential for planning, especially in the energy, finance, health and climate sectors. While recent deep learning models show remarkable improvement over statistical methods, many models, such as iTransformer and NLinear, do not explicitly incorporate periodicity, a defining feature of time series data. To address this, we propose two enhanced models: <b>FANLinear</b> and <b>iFANformer</b>, which integrate the Fourier Analysis Network (FAN) in the architecture of NLinear and iTransformer, respectively. We evaluate our models on five benchmark datasets (ECL, ETT-h1, ETT-h2, Traffic, and Weather) across multiple forecast horizons. Our method reduces the average Mean Squared Error (MSE) by 12.53% for iFANformer and 24.16% for FANLinear, when compared to base models. Our results show that explicitly adding periodicity modeling can enhance the performance of base models, enabling them to compete with and even outperform state-of-the-art (SOTA) models.</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

Enhancing time series forecasting by learnable fourier-driven periodicity

  • 期刊论文

  • 发表日期：

    2026年3月6日

查看全文
  • Tayyab Saeed Qureshi,
  • Asma Ahmad Farhan

摘要

Long-horizon Time Series Forecasting (LTSF) is essential for planning, especially in the energy, finance, health and climate sectors. While recent deep learning models show remarkable improvement over statistical methods, many models, such as iTransformer and NLinear, do not explicitly incorporate periodicity, a defining feature of time series data. To address this, we propose two enhanced models: FANLinear and iFANformer, which integrate the Fourier Analysis Network (FAN) in the architecture of NLinear and iTransformer, respectively. We evaluate our models on five benchmark datasets (ECL, ETT-h1, ETT-h2, Traffic, and Weather) across multiple forecast horizons. Our method reduces the average Mean Squared Error (MSE) by 12.53% for iFANformer and 24.16% for FANLinear, when compared to base models. Our results show that explicitly adding periodicity modeling can enhance the performance of base models, enabling them to compete with and even outperform state-of-the-art (SOTA) models.