TFDNet: Time-Frequency Enhanced Decomposed Network for Long-term Time Series Forecasting

• DOI: 10.48550/arxiv.2308.13386
• Publication Date: 2023-01-01
• Authors: Yuxiao Luo; Ziyu Lyu; Xingyu Huang
• Keywords: Benchmark (surveying); Kernel (algebra)

Long-term time series forecasting is a vital task and has wide range of real applications. Recent methods focus on capturing the underlying patterns from one single domain (e.g. or frequency domain), have not taken holistic view to process long-term time-frequency domains. In this paper, we propose Time-Frequency Enhanced Decomposed Network (TFDNet) capture both temporal periodicity domain. TFDNet, devise multi-scale enhanced encoder backbone develop two separate trend seasonal blocks distinct within decomposed components in multi-resolutions. Diverse kernel learning strategies operations been explored, by investigating incorporating potential different channel-wise correlation multivariate series. Experimental evaluation eight datasets five benchmark domains demonstrated that TFDNet superior state-of-the-art approaches effectiveness efficiency.