MSFI: Multi-Scale Frequency Interpolation for Time Series Forecasting
Jihao Zhang, Zhijiang Wang, Chunna Zhao · 2024
Time series forecasting is employed in various fields such as electricity and transportation. However, complex temporal changes in data remain a challenging task for long-term forecasting. From a multi-scale perspective, it can be observed that the time series exhibits different states at different sampling scales. Fine-grained information and coarse-grained information are reflected at small and large scales, respectively. To address this, we propose a new Multi-Scale Frequency Interpolation model for time series forecasting. This model consists of a Multi-Scale (MS) architecture and a Frequency Interpolation (FI) module. Specifically, MS architecture decomposes time series from multiple views to increase the model’s ability to perceive at different scales. And FI module further performs frequency decomposition and interpolation prediction on different scales. Thus, the complex temporal variations of the time series are internally unraveled. Extensive experiments on 5 real world datasets demonstrate the superiority of our approach.