面向预测稳定性的通道感知时频解耦多变量时序预测

Channel-aware time–frequency disentanglement for multivariate time series forecasting under complex dependency structures

摘要

多变量时序预测是智能调度、风险防控与资源配置的核心技术. 尽管 Transformer 模型特征表达能力强大，但在通道建模与监督机制上仍存在不足. 在高维场景下，直接对所有变量统一建模会引入大量冗余与噪声通道，导致通道结构的失真. 这不仅干扰了关键时序特征的提取，更会在多步迭代预测中破坏模型对标签自回归依赖的建模，引发预测误差的级联放大，极大地加剧了监督偏差与预测不稳定性. 为此，本文提出一种面向预测稳定性的通道感知时频解耦预测模型（CATFD）.该模型采用主辅双分支结构：主分支融合混合专家机制与多尺度解耦，挖掘多尺度时序特征. 辅分支构建可学习通道掩码，在频域中筛除冗余通道，并通过掩码注意力机制调控主分支建模路径. 在此基础上，引入双域动态联合监督机制协同优化时频域损失，并显式建模标签自相关结构，有效缓解监督偏差问题. 实验结果表明，CATFD 模型在电力、交通、金融等多个真实数据集上均取得了优异表现，不仅验证了其有效性与泛化性，同时也展示了在噪声环境下的鲁棒性与抗扰动能力.

Abstract

Multivariate time-series forecasting is a core technology for a wide range of critical applications, including intelligent scheduling, risk management, and resource allocation in complex modern systems. Although transformer-based models have demonstrated formidable capabilities in capturing long-range dependencies through sophisticated self-attention mechanisms, their practical effectiveness in high-dimensional scenarios remains significantly restricted by inherent deficiencies in channel modeling strategies and supervision mechanisms. In typical high-dimensional real-world settings, modeling all variables jointly and uniformly tends to introduce a disproportionate number of redundant or weakly correlated channels, which inevitably leads to distortion of the underlying channel structures. This structural distortion not only interferes with the extraction of critical informative temporal features, but also disrupts the model's ability to accurately capture and learn the label autoregressive dependencies inherently prevalent in future sequences, triggering a cascading amplification of prediction errors that exacerbates supervision bias and degrades forecasting stability. To address these multifaceted challenges, this paper proposes a novel channel-aware time–frequency disentangled (CATFD) forecasting model specifically oriented toward achieving long-term prediction stability. The model adopts a sophisticated dual-branch architecture comprising a main branch and an auxiliary branch to synergistically refine the temporal dynamics and purify channel structures. The main branch focuses on deep temporal modeling by integrating a mixture-of-experts mechanism with a multiscale decoupling module, where a dynamic routing system enables the adaptive selection of specialized representations under heterogeneous patterns, whereas adaptive filters capture multiscale dynamics by effectively separating long-term trends from seasonal fluctuations. Simultaneously, the auxiliary branch operates within the frequency domain to construct a learnable channel mask using a soft sparsity mechanism designed to filter out noisy and redundant channels, thereby ensuring that the model concentrates on the most informative variables. To achieve seamless cross-branch synergy, a masked attention mechanism is introduced to dynamically regulate the modeling paths of the main branch based on the refined structural importance provided by the auxiliary branch. Consequently, we further introduce a dual-domain dynamic joint supervision mechanism to collaboratively optimize time-domain prediction losses and frequency-domain reconstruction losses. By explicitly incorporating label autoregressive dependencies and leveraging the orthogonal properties of the frequency domain to weaken interlabel interference, this approach effectively mitigates the supervision bias problem and ensures temporal consistency across the prediction steps. Extensive experiments are conducted on multiple real-world datasets from diverse domains, including power systems (Electricity transformer temperature, ETT), transportation networks (Traffic), and financial markets (Exchange). The experimental results demonstrate that CATFD consistently achieves superior performance compared with representative state-of-the-art transformer- and graph-based baselines, with gains being particularly evident in long-horizon tasks, where error accumulation is traditionally the most significant. Furthermore, robustness tests conducted under various noisy environments and perturbation conditions verify that CATFD maintains stable performance with limited degradation, showcasing its exceptional anti-interference capability. These results collectively indicate that the proposed method significantly improves prediction consistency and generalization ability, providing a robust solution for multivariate time-series forecasting in complex and uncertain real-world scenarios.

关键词

多变量时序预测 / 通道感知建模 / 时频解耦 / 自回归依赖 / 监督偏差

Key words

multivariate time series forecasting / channel-aware modeling / time–frequency disentanglement / autoregressive dependency / supervision bias

参考文献

[1] Chen Z L, Wu Z H, Cheung W K, et al. Hydrothermal preparation and microstructure analysis of silver tin oxide contact materials[J]. Chin J Eng, 2007, 29(10): 1023

[2] Liu Q H, Boniol P, Palpanas T, et al. Time—series anomaly detection: Overview and new trends[J]. Proc VLDB Endow, 2024, 17(12): 4229

[3] Fan S M, Wang H, Zhang F. CAWformer: A cross variable attention with discrete wavelet denoising for multivariate time series forecasting[J]. Knowl Based Syst, 2025, 324: 113846

[4] Stein G, Shadaydeh M, Blunk J, et al. CausalRivers: Scaling up benchmarking of causal discovery for real—world time—series[C]// The Thirteenth International Conference on Learning Representations , 2025: 92778

[5] Dang Y Z, Yang E N, Guo G B, et al. Uniform sequence better: Time interval aware data augmentation for sequential recommendation[C]// Proceedings of the AAAI Conference on Artificial Intelligence, 2023: 4225

[6] Yang J, Zhang K, Zhang G, et al. Glocal information bottleneck for time series imputation[C]// The Thirty—ninth Annual Conference on Neural Information Processing Systems, 2025: 104452

[7] Wang H, Chen Z C, Liu Z R, et al. Entire space counterfactual learning for reliable content recommendations[J]. IEEE Trans Inf Forensics Secur, 2025, 20: 1755

[8] Vaswani A, Shazeer N, Parmar N, et al. Attention is all you need[C]// Advances in Neural Information Processing Systems 30, 2017: 5998

[9] Liu Y, Hu T G, Zhang H R, et al. iTransformer: Inverted transformers are effective for time series forecasting[PP/OL]. arXiv (2023—10—10) [2026—01—26]. https://arxiv.org/abs/2310.06625

[10] Nie Y Q, Nguyen N H, Sinthong P, et al. A time series is worth 64 words: Long—term forecasting with transformers[PP/OL]. arXiv (2022—11—27) [2026—01—26]. https://arxiv.org/abs/2211.14730

[11] Zhang Y H, Yan J C. Crossformer: Transformer utilizing cross—dimension dependency for multivariate time series forecasting[C]// The Eleventh International Conference on Learning Representations, 2023: 1

[12] Liu Q, Fang Y, Jiang P, et al. DGCformer: Deep graph clustering transformer for multivariate time series forecasting[PP/OL]. arXiv (2024—05—14) [2026—01—26]. https://arxiv.org/abs/2405.08440

[13] Lin S S, Lin W W, Wu W T, et al. SparseTSF: Modeling long—term time series forecasting with 1k parameters[C]// Proceedings of the 41st International Conference on Machine Learning, 2024: 30211

[14] Wang Y, Qiu Y, Chen P, et al. LightGTS: A lightweight general time series forecasting model[C]// Forty—second International Conference on Machine Learning, 2025: 64109

[15] Chen S A, Li C L, Arik S O, et al. TSMixer: An all—MLP architecture for time series forecasting[PP/OL]. arXiv (2023—03—10) [2026—01—26]. https://arxiv.org/abs/2303.06053

[16] Zhou H Y, Zhang S H, Peng J Q, et al. Informer: Beyond efficient transformer for long sequence time—series forecasting[C]// Proceedings of the AAAI Conference on Artificial Intelligence. 2021: 11106

[17] Wu H X, Xu J H, Wang J M, et al. Autoformer: Decomposition transformers with auto—correlation for long—term series forecasting[C]// Advances in Neural Information Processing Systems 34. 2021: 22419

[18] Shih S Y, Sun F K, Lee H Y. Temporal pattern attention for multivariate time series forecasting[J]. Mach Learn, 2019, 108(8—9): 1421

[19] Cini A, Zambon D, Alippi C. Sparse graph learning from spatiotemporal time series[J]. J Mach Learn Res, 2023, 24(242): 1

[20] Louizos C, Welling M, Kingma D P. Learning sparse neural networks through L₀ regularization[PP/OL]. arXiv (2017—12—04) [2026—01—26]. https://arxiv.org/abs/1712.01312

[21] Wang Z W, Lu J W, Tao C X, et al. Learning channel—wise interactions for binary convolutional neural networks[C]// 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019: 568

[22] Zhang K, Guo Y R, Wang X S, et al. Channel—wise and feature—points reweights densenet for image classification[C]// 2019 IEEE International Conference on Image Processing (ICIP), 2019: 410

[23] Wang H, Pan L C, Shen Y, et al. Fredf: Learning to forecast in the frequency domain[C]// The Thirteenth International Conference on Learning Representations, 2025: 1

[24] Zhou T, Ma Z Q, Wang X, et al. Fedformer: Frequency enhanced decomposed transformer for long—term series forecasting[C]// International Conference on Machine Learning. PMLR, 2022: 27268

[25] Yi K, Zhang Q, Fan W, et al. Frequency—domain MLPs are more effective learners in time series forecasting[C]// Advances in Neural Information Processing Systems 36, 2023: 76656

[26] Cao D F, Wang Y J, Duan J Y, et al. Spectral temporal graph neural network for multivariate time—series forecasting[C]// Advances in Neural Information Processing Systems 33, 2020: 17766

[27] Liu Z, Luo Y C, Li B Y, et al. Learning soft sparse shapes for efficient time—series classification[C]// Forty—second International Conference on Machine Learning, 2025: 39032

[28] Cheng Y. Multivariate time series forecasting through automated feature extraction and transformer—based modeling[J]. J Compute Sci Soft App, 2025, 5(5). https://doi.org/10.5281/zenodo.15381947

[29] Bandara K, Hyndman R J, Bergmeir C. MSTL: A seasonal—trend decomposition algorithm for time series with multiple seasonal patterns[J]. Int J Oper Res, 2025, 52(1): 79

[30] Hu Y F, Liu P Y, Zhu P, et al. Adaptive multi—scale decomposition framework for time series forecasting[C]// Proceedings of the AAAI Conference on Artificial Intelligence. 2025: 17359

[31] Goldberger J, Hinton G E, Roweis S, et al. Neighbourhood components analysis[C]// Advances in Neural Information Processing Systems 17, 2004: 513

[32] Dai T, Wu B L, Liu P Y, et al. Periodicity decoupling framework for long—term series forecasting[C]// The Twelfth International Conference on Learning Representations, 2024

[33] Peng C, Zhang Y, Cheng Y, et al. Pathformer: Multi—scale transformers with adaptive pathways for time series forecasting[C]// The Twelfth International Conference on Learning Representations, 2024

[34] Liu Y, Wu H X, Wang J M, et al. Non—stationary transformers: Exploring the stationarity in time series forecasting[C]// Advances in Neural Information Processing Systems 35, 2022: 9881

[35] Wang S, Li J, Shi X, et al. TimeMixer++: A general time series pattern machine for universal predictive analysis[C]// The Thirteenth International Conference on Learning Representations, 2025

基金资助

国家重点研发国际合作项目(2023YFE0112500)

河南省重大产业关键技术攻关“揭榜挂帅”项目(251000210300)

河南省重点研发专项(241111211200)

河南省重点研发专项(251111211100)

河南省科技攻关项目(262102211001)