UNetPlusTS: Decomposition-Mixing UNet++ Architecture for Long-Term Time Series Forecasting

Long-term time series forecasting has widespread applications in real life, including finance, traffic, weather and sensor data analysis. Time series possess seasonal, trend and irregular components. However, current MLP-based mixing models fall short in effectively capturing multi-scale sequential information and achieving comprehensive feature fusion. The sequence information becomes non-stationary, and as the network depth increases, issues with missing or distorted information flow may arise. In our study, we propose a novel architecture called UNetPlusTS that combines the strengths of Mixer models and Linear models to enhance forecasting capabilities. Specifically, we split series into multiple channels, apply seasonal-trend decomposition to each series and process them independently using our meticulously designed UNet ++ architecture named UNetPlus Mixing Module (UPM). Combined with our unique sampling strategy, it promotes deep integration of seasonal and trend components, thereby alleviating non-stationarity. Experimental results on multiple real-world datasets show that UNetPlusTS has significantly improved the forecasting accuracy, demonstrating its effectiveness and robustness.