Fourier U-Shaped Network for Multi-Variate Time Series Forecasting.

Multi-variate time series forecasting plays a crucial role in addressing key tasks across various domains, such as early warning, pre-planning, resource scheduling, and other critical tasks. Thus, accurate multi-variate time series forecasting is of significant importance in guiding practical applications and facilitating these essential tasks. Recently, Transformer-based multi-variate time series forecasting models have demonstrated tremendous potential due to their outstanding performance in long-term time predictions. However, Transformer-based models for multi-variate time series forecasting often come with high time complexity and computational costs. Therefore, we propose a low time complexity model called Fourier U-shaped Network (F-UNet) for multi-variate time series forecasting, which is non-Transformer based. Specifically, F-UNet is composed of low time complexity neural network components, such as Fourier neural operator and feed-forward neural network, arranged in a U-shaped architecture. F-UNet conducts channel and temporal modeling separately for the multi-variate time series. The U-Net constructed based on Fourier neural operators is employed to achieve channel interactions, while linear layers are used to realize temporal interactions. Experimental results on several real-world datasets demonstrate that F-UNet outperforms existing Transformer-based models with higher efficiency in multi-variate time series forecasting.