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.