S2LNet: Review the Non-Stationarity in Multivariate Time Series Forecasting

Transformer-based methods have achieved remarkable advances in multivariate time series forecasting for their long-range ability. However, the non-stationarity of real-world time series, make these models particularly prone to overfitting when data distribution changes over time. Recently, despite various attempts in existing studies, they either overlook cross-channel mutual information gains or struggle to effectively capture cross-time features. To overcome these limitations, we review the characteristics of time series and develop a novel Short-term to Long-term network called S2LNet, which combines short-term cross-time features into long-term distributions and then models cross-channel dependencies models cross-time and cross-channel dependencies. For cross-time features, S2LNet first decomposes the input sequence into seasonal and trend items, then employs Transformers for capturing seasonal features seasonal items and multilayer perceptrons MLPs for trend items modeling trend features. These modeled short-term features are then fused and downsampled into long-term relationships through the Long-term Fusion module, followed by a channel-wise Transformer for long-term cointegration across channels. Extensive experiments on various real-world benchmarks have verified the superiority of our model over other state-of-the-art baselines.