This paper presents a novel method to address the limitations of fixed-length inputs in long-term time series forecasting. Existing deep learning models suffer from overfitting as input lengths increase, leading to a rapid decline in accuracy. This is attributed to the combination of multi-scale patterns in time series and the fixed focus scale of the model. This study finds that patterns appearing at various scales in time series reflect multi-cycle characteristics, with each scale corresponding to a specific cycle length. Furthermore, it reveals that token size determines the model's behavior, determining the scales it focuses on and the context size it can accommodate. Therefore, we propose a novel method that separates multi-scale temporal patterns in time series and models each pattern as a token size representing its corresponding cycle length. By introducing the Time Series Decomposition Module (MPSD) and the Multi-Token Pattern Recognition Neural Network (MTPR), we can process inputs up to 10x longer, improving performance by up to 38% while reducing computational complexity by 0.22x and enhancing interpretability.
