FTLinear: MLP based on Fourier Transform for Multivariate Time-series Forecasting
Listed in
This article is not in any list yet, why not save it to one of your lists.Abstract
Most existing multivariate time series forecasting models are based on Recurrent Neural Networks(RNNs), Graph Neural Networks(GNNs), and Transformers, which leads to high model complexity. Although linear models based on Multi-layer Perceptrons(MLPs) have the advantages of low complexity and high performance, they fail to capture the long-term dependencies in time series, resulting in poor forecasting performance. This paper proposes a linear model based on Fourier Transform, FTLinear, which first converts the time series to the frequency domain, then uses frequency enhancement to remove noise from the data, and finally obtains the output through a simple linear transformation. Experimental results show that the proposed model improves MSE and MAE by 18.48% and 8.03% respectively compared to iTransformer, by 12.03% and 2.35% respectively compared to PatchTST, and by 17.75% and 7.52% respectively compared to DLinear on 8 public datasets. This demonstrates that converting time series to the frequency domain can provide the model with a global perspective, enhancing its ability to capture long-term dependencies and thus improving forecasting performance.