CFANet: Channel-Frequency Adaptive Network for Long-Term Time Series Forecasting

Abstract

Frequency domain modeling methods have emerged as a promising research direction for long-term series forecasting, owing to their natural advantages in capturing global dependencies and reducing computational complexity. However, existing methods overlook the heterogeneity and dynamic characteristics of frequency properties across different channels. These methods adopt a unified frequency-domain processing strategy and static filtering approaches for all channels, which limits the model’s ability to capture long-term temporal dependencies. To address this issue, this paper proposes a Channel Frequency Adaptive Network (CFANet) for long-term time series forecasting. CFANet designs a dual adaptive mechanism: (i) a channel adaptive mechanism that enables each channel to learn optimal frequency response patterns through channel-specific frequency filters; (ii) a frequency adaptive mechanism that allows filters to dynamically adjust frequency weights through end-to-end learning, thereby adapting to evolving frequency patterns. Unlike existing models with channel-shared and static filtering strategies, CFANet achieves channel-level adaptive enhancement in the frequency domain. In extensive experiments on nine benchmark datasets encompassing 72 test scenarios, CFANet achieves optimal performance in 50 scenarios, reducing prediction errors by 9.8% on average compared to state-of-the-art baselines, demonstrating consistently superior performance across diverse forecasting tasks. Notably, on heterogeneous datasets where channel spectral characteristics diverge significantly, CFANet delivers 10%–15% error reductions, confirming the effectiveness of the proposed channel frequency adaptive mechanism.