.png){kind=icon}
.png){kind=icon}
Return
Super-Long Input Sequences for Long-Term Time Series Forecasting With Missing Values
.png){kind=icon}
PRE
AI
DOI：10.1109/TIM.2025.3627351.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
The time series data collected by instruments and sensors are critical for monitoring modern systems and supporting informed decision-making. However, the real-world time series often contain missing values, which significantly hinder the performance of forecasting models. Existing methods for addressing this challenge are limited and typically rely on additional imputation techniques, which risk distorting valid data and increasing the computational cost. In this work, SLNet, which is a novel model for robust long-term time series forecasting (TSF) under conditions of missing values, is presented. SLNet addresses missing values by leveraging super-long input sequences (SLISs) and incorporating all available historical observations for arbitrary prediction horizons. A focal encoder is employed to effectively and adaptively capture both short-term and long-term dependencies, which enables the extraction of comprehensive temporal features. Furthermore, SLNet produces predictions that are accompanied by confidence scores, which are generated by tokenizing the output on the basis of local sequence characteristics, thereby increasing accuracy and enhancing interpretability. Extensive experiments on four public benchmarks demonstrate that SLNet achieves state-of-the-art performance in both conventional long-term forecasting tasks and more challenging scenarios that involve missing values. The code is released on https://github.com/OrigamiSL/SLNet
Keywords:
Missing data
super-long input sequence
time series forecasting (TSF)
tokenization
Journal
IF:
5.9
Papers:
1.9W
Citations:
5.8W