Service Oriented Computing and Applications Information Systems
A dual-branch time series forecasting approach based on frequency decomposition
Abstract
Real-world time series data is often highly non-stationary, typically exhibiting low-frequency dominated long-term trends and periodic structures alongside high-frequency short-term fluctuations and noise. These distinct frequency components are highly coupled within the time domain. Existing time-domain decomposition methods typically use fixed-size smoothing windows for trend extraction. This approach often blurs information or attenuates local dynamic features, making precise frequency isolation difficult. Furthermore, it is challenging for a single model to simultaneously balance long-range dependency modeling and local detail modeling, thereby limiting the accuracy of long-term forecasting. To address these issues, this paper proposes a dual-branch time series forecasting approach based on frequency decomposition. First, Reversible Instance Normalization (RevIN) is applied to mitigate the non-stationary statistical distribution shifts in the data. Subsequently, Fast Fourier Transform (FFT) is utilized to decompose the sequence in the frequency domain before transforming it back into the time domain. During the forecasting phase, a dual-branch architecture is designed: a Low-frequency Patch-based Forecasting Block (PFB) processes the low-frequency components, while a High-frequency Convolutional Forecasting Block (CFB) handles the high-frequency components. This design not only captures long-term trends and periodic structures via the patch mechanism and self-attention, but also fully leverages dilated causal convolutions to characterize local abrupt changes and short-term volatility. Finally, the prediction results from both branches are fused and denormalized to yield the final output sequence. Experimental results on seven benchmark datasets demonstrate that the proposed method achieves significant performance improvements in long-term time series forecasting tasks.