IFTNet: Interpolation Frequency- and Time-Domain Network for Long-Term Time Series Forecasting

Long-term time series forecasting has widespread applications in multiple fields. Time series possess seasonality and trend, and existing models predict time series in either the time or frequency domain, with the former generally struggling to learn seasonality and the latter overlooking trend. Additionally, the Picket Fence Effect leads to bias in frequency domain observation, resulting in inaccurate capturing of frequency domain features. In our study, we propose a framework called Interpolation Frequency- and Time-domain Network (IFTNet) for time series forecasting, which considers both the frequency and time domains, as well as seasonality and trend of time series. Specifically, we model the seasonal and trend components of time series separately. We introduce a novel module for frequency-domain interpolation to mitigate the impact of the Picket Fence Effect. We also design a multi-scale depth-wise attention module to capture the features at different scales within and during the time series periods. The experimental results show that compared to other models, IFTNet has a stable improvement in prediction accuracy across multiple real-time series datasets.