Online Time Series Prediction via Self-Supervised Learning and Data Replay
Tingting Zhang, Bo Ding, Anyi Li, Jing Mei Yang · 2025
Online time series prediction plays a crucial role in achieving accurate prediction and dynamic adjustment. Despite its effectiveness, online time series forecasting faces two key challenges. On the one hand, influenced by the generation mechanism of real-time data, online time series forecasting often encounters the issue of insufficient data, which prevents the model from effectively representing the intrinsic patterns of the data. Consequently, the model tends to underfit, which negatively impacts its predictive performance. On the other hand, the model tends to fit the most recent data, which may focus excessively on short-term fluctuations while ignoring long-term trends and overall patterns. This tendency may cause distribution shift, which limits the model's generalization. To address these challenges, we propose Self-Supervised learning and data Replay for Online Time Series prediction (SSR4OTS). Our method adopts a self-supervised learning strategy to improve the model's predictive performance by increasing the utilization of available data through data augmentation and self-distillation mechanism. Furthermore, SSR4OTS uses a data replay approach to allow the model to learn more comprehensive patterns to enhance the stability of predictions. Experimental results demonstrate that SSR4OTS significantly improves accuracy and reliability of online time series forecasting.