We found a match
Your institution may have access to this item. Find your institution then sign in to continue.
- Title
Temporal patterns decomposition and Legendre projection for long-term time series forecasting.
- Authors
Liu, Jianxin; Ma, Tinghuai; Su, Yuming; Rong, Huan; Khalil, Alaa Abd El-Raouf Mohamed; Wahab, Mohamed Magdy Abdel; Osibo, Benjamin Kwapong
- Abstract
Long-term time series forecasting (LTSF) means utilizing historical data to forecast future sequences that are relatively distant in time, providing support for long-term warnings, planning, and decision-making. LTSF is more challenging than short-term forecasting due to its larger output length. It requires forecasting methods to accurately capture long-term temporal dependencies from complex sequences with intertwined temporal patterns. For LTSF tasks, existing works propose variants of recurrent neural networks, convolutional neural networks, and transformers to catch temporal dependencies. However, these methods usually suffer from the insufficient ability to capture long-term temporal dependencies and excessively high complexity, resulting in unreliable forecasting performance. Therefore, we propose an LTSF method based on temporal patterns decomposition and Legendre projection (TPDLP). Firstly, we use temporal patterns decomposition to handle complex temporal patterns to perform decomposed refinement forecasting. Subsequently, we use high-order Legendre polynomial projection with a signal transfer module based on multilayer perceptron networks to capture long-term temporal dependencies, thereby achieving LTSF. Furthermore, we introduce targeted data normalization to alleviate the impact of distribution shifts on sequence forecasting. Through extensive experimentation with six popular real-world datasets, our TPDLP model shows an average relative improvement of 15.8% compared to the best baseline in terms of performance, measured by prediction error. In addition, it also demonstrates superior efficiency, which showcases its utility in real-world applications. Code is available at this repository: https://github.com/JoeDoex/TPDLP.
- Subjects
CONVOLUTIONAL neural networks; TRANSFORMER models; TIME-varying networks; TIME series analysis; FORECASTING
- Publication
Journal of Supercomputing, 2024, Vol 80, Issue 16, p23407
- ISSN
0920-8542
- Publication type
Article
- DOI
10.1007/s11227-024-06313-4