FFTNet: Fusing Frequency and Temporal Awareness in Long-Term Time Series Forecasting

Time series forecasting is extensively utilised in meteorology, transportation, finance, and industrial domains. Precisely recognising cyclical trends and abrupt local changes in time series is essential for enhancing forecasting accuracy. Frequency-domain representations are adept at identifying periodic traits, whereas time-domain approaches are superior for spotting localised quick changes. Traditional techniques sometimes prioritise a single domain, overlooking the advantages of integration. This paper introduces a novel hybrid model, FFTNet, that simultaneously pulls characteristics from both domains to optimise their respective benefits. Theoretical examination indicates that the 2D CNN utilises dual-axis convolution kernels to jointly describe global cross-patch structures and local temporal patterns, whilst the frequency-domain MLP enhances spectral components in accordance with Parseval’s Theorem and the Convolution Theorem. A frequency-domain MLP is employed to discern periodic and trend characteristics, while a 2D CNN in the time domain identifies localised abrupt changes. This hybrid methodology differentiates itself from previous methods that depend exclusively on a single domain, providing a more thorough comprehension of the underlying patterns in time series data. Experiments on seven real-world datasets indicate that FFTNet surpasses existing techniques, attaining state-of-the-art performance with enhancements of 11.8% and 4.7% in MSE and MAE, respectively.