Time Series Forecasting Based on a Multi-Frequency Fully Actuated System
Zhiyuan Liu, Liang Zhao, Rui Lin, Kai Yan, Yi Qing Yang · 2025
Time series forecasting has a wide range of applications in real-world scenarios. Recent advances in this field have leveraged frequency-domain decomposition techniques to transform data from the time domain to the frequency domain for more effective modeling. However, components with different frequency bands often correspond to distinct underlying patterns, indicating that they follow different dynamics. To better differentiate and model these components, we propose a novel forecasting model based on the Fully Actuated Time Series system (FATS), inspired by control theory. FATS establishes a fully actuated system across multiple frequency bands, enabling independent modeling of the components corresponding to different frequency ranges. This design not only enhances the model's forecasting capability but also significantly reduces the number of model parameters. Furthermore, we introduce an independent modeling strategy for the real and imaginary parts of complex frequency-domain data, allowing for more precise representation and processing of such information. Extensive experiments conducted on multiple datasets across various domains demonstrate that our proposed model achieves state-of-the-art performance.