UnetTSF: A Better Performance Linear Complexity Time Series Prediction Model

TL;DR

This work addresses the efficiency-accuracy trade-off in long-horizon time series forecasting by introducing UnetTSF, a linear-complexity predictor that replaces binary trend-season decomposition with a time-series feature pyramid (FPN) and a U‑Net–inspired encoder–decoder. The method combines a pooling-based multilevel feature extractor with a multi-stage fusion mechanism to produce forecasts of length $T$ with complexity $O(L)$. Empirically, UnetTSF achieves substantial improvements over the linear baseline DLlinear (average MSE down by 10.1% and MAE down by 9.1%) and competitive results versus PatchTST across eight datasets, while drastically reducing computational cost (MACs) compared with Transformer-based models. The approach demonstrates strong potential for production deployment in resource-limited settings due to its favorable accuracy-efficiency balance and robust multivariate performance.

Abstract

Recently, Transformer-base models have made significant progress in the field of time series prediction which have achieved good results and become baseline models beyond Dlinear. The paper proposes an U-Net time series prediction model (UnetTSF) with linear complexity, which adopts the U-Net architecture. We are the first to use FPN technology to extract features from time series data, replacing the method of decomposing time series data into trend and seasonal terms, while designing a fusion structure suitable for time series data. After testing on 8 open-source datasets, compared to the best linear model DLiner. Out of 32 testing projects, 31 achieved the best results. The average decrease in mse is 10.1%, while the average decrease in mae is 9.1%. Compared with the complex transformer-base PatchTST, UnetTSF obtained 9 optimal results for mse and 15 optimal results for mae in 32 testing projects.

Figures (2)

• Figure 1: The pipeline of existing Model TSF solutions.
• Figure 2: UnetTSF architecture.(a)The overall inference process of UnetTSF model.The model consists of downsampling and fully connected components, with linear complexity.(b)Using avgpool to complete FPN sampling of temporal data.(c)Illustration of the basic linear model.