U-Mixer: An Unet-Mixer Architecture with Stationarity Correction for Time Series Forecasting

TL;DR

This paper tackles non-stationarity in time series forecasting by introducing U-Mixer, a Unet-Mixer architecture that processes time-series patches while separating temporal and channel interactions. A novel stationarity correction method aligns distributional changes via an autocorrelation constraint using an affine scale α computed through FFT-based Wiener-Khinchin acceleration, updating predictions with a mean adjustment term. The approach combines per-patch embeddings, a Unet encoder-decoder, and an MLP-based mixer to capture multi-level local patterns, achieving state-of-the-art performance on six real-world datasets for both long- and short-term forecasting. The results demonstrate robust improvements over SOTA baselines and highlight the method's effectiveness in handling non-stationary regimes with strong practical impact for real-world forecasting tasks.

Abstract

Time series forecasting is a crucial task in various domains. Caused by factors such as trends, seasonality, or irregular fluctuations, time series often exhibits non-stationary. It obstructs stable feature propagation through deep layers, disrupts feature distributions, and complicates learning data distribution changes. As a result, many existing models struggle to capture the underlying patterns, leading to degraded forecasting performance. In this study, we tackle the challenge of non-stationarity in time series forecasting with our proposed framework called U-Mixer. By combining Unet and Mixer, U-Mixer effectively captures local temporal dependencies between different patches and channels separately to avoid the influence of distribution variations among channels, and merge low- and high-levels features to obtain comprehensive data representations. The key contribution is a novel stationarity correction method, explicitly restoring data distribution by constraining the difference in stationarity between the data before and after model processing to restore the non-stationarity information, while ensuring the temporal dependencies are preserved. Through extensive experiments on various real-world time series datasets, U-Mixer demonstrates its effectiveness and robustness, and achieves 14.5\% and 7.7\% improvements over state-of-the-art (SOTA) methods.

Figures (5)

• Figure 1: The architecture of U-Mixer, which consists of per-patch embeddings, Unet encoder-decoder, stationarity correction and a forecasting head.
• Figure 2: The Unet encoder-decoder of U-Mixer. Encoders and decoders are both MLP blocks. The term "Merge" refers to the combining process of features from different levels.
• Figure 3: The MLP block. (a) is the MLP block, which contains one temporal MLP layer and one channel MLP layer. (b) is the specific structure of MLP, which consists of two fully-connected layers and a GELU nonlinearity.
• Figure 4: Visualization of forecasting results on multiple datasets by U-Mixer.
• Figure 5: Performance comparison on varying.