SDHNet: a sampling-based dual-stream hybrid network for long-term time series forecasting

Recently, deep learning models have achieved notable success in long-term time series forecasting. However, real-world time series data typically exhibit complex temporal patterns, characterized by both short-term and long-term variations across multiple time scales. This complexity makes it difficu...

Full description

Saved in:
Bibliographic Details
Published in:The Journal of supercomputing 2025, Vol.81 (1), Article 68
Main Authors: Ma, Shichao, Miao, Shengfa, Yao, Shaowen, Jin, Xin, Chu, Xing, Yu, Qian, Tian, Yuling, Wang, Ruoshu
Format: Article
Language:English
Subjects:
Compilers
Complexity
Computer Science
Fast Fourier transformations
Forecasting
Fourier series
Fourier transforms
Interpreters
Processor Architectures
Programming Languages
Representations
Sampling
Time series
Citations: Items that this one cites
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Recently, deep learning models have achieved notable success in long-term time series forecasting. However, real-world time series data typically exhibit complex temporal patterns, characterized by both short-term and long-term variations across multiple time scales. This complexity makes it difficult to effectively distinguish and integrate these different patterns. To address this challenge, we propose a Sampling-based Dual-stream Hybrid Network ( SDHNet ), designed specifically to disentangle short-term and long-term variations inherent in one-dimensional (1D) time series data. The core mechanism of SDHNet involves applying continuous and equidistant periodic sampling strategies based on fast Fourier transform (FFT) to generate short-term and long-term representations in a two-dimensional (2D) space. The short-term representations are optimized for capturing localized, high-frequency patterns, while the long-term representations are crucial for identifying global dependencies and trends. To fully extract this information, SDHNet adopts a dual-stream framework, modeling both types of representations in parallel, rather than using a conventional sequential architecture. In addition to its effectiveness, SDHNet demonstrates greater efficiency with longer sequence inputs and shorter inference times. Extensive experiments show that SDHNet consistently outperforms established baseline models in both multivariate and univariate time series forecasting tasks. Our code is accessible at this repository: https://github.com/Renaissance5/SDHNet .
ISSN:0920-8542
1573-0484
DOI:10.1007/s11227-024-06495-x