FAformer: Frequency Analysis Transformer for Multivariate Time Series Forecasting

Time series forecasting plays a key role in many fields such as energy, transport and industrial production. In recent years, Transformer-based models have achieved impressive results in the field of time series forecasting, demonstrating strong modelling capabilities and generalization performance. However, computing sequence correlation through the attention mechanism has quadratic complexity and limited ability to capture subtle features in the data and prevent noise interference. In order to overcome these challenges, this paper proposes a new frequency analysis based Transformer framework, named FAformer. In FAformer, in order to fully exploit the latent information in the time series, we employ a multilevel discrete wavelet transform to decompose the series into signals at different frequencies, and then design two independent modules for approximation and detail components to capture different latent patterns at multiple resolutions. In addition, we employ the Fourier transform to remove noise and utilize a window attention mechanism to reduce model complexity. The experimental results demonstrate the effectiveness of F Aformer in terms of prediction capacity.