Channel-Masked Time-Frequency Disentanglement for Multivariate Time Series Forecasting

Multivariate time series forecasting (MTSF) is a pivotal technique for intelligent scheduling, risk control, and resource allocation. Despite their powerful representational capacity, Transformer-based models often struggle with two key challenges in MTSF: ineffective channel modeling due to redundant or weakly correlated variables, and supervision bias stemming from the neglect of label auto-correlation. To overcome these challenges, we propose the Channel-Masked Time-Frequency Disentangled model (CM-TFD). Our framework features a primary-auxiliary dual-branch architecture that synergistically models temporal patterns and channel relationships. The primary branch integrates a Mixture of Experts mechanism with a time-frequency cascading module to capture multi-scale temporal features. Simultaneously, the auxiliary branch constructs a learnable channel mask in the frequency domain to filter out redundant information and dynamically guides the primary branch via a soft sparsity mechanism. Furthermore, we introduce a cross-domain dynamic weight adjustment mechanism to jointly optimize time-domain and frequency-domain loss functions. By explicitly modeling the label auto-correlation structure, our method effectively mitigates supervision bias. Extensive experiments on nine real-world datasets from diverse domains (e.g., electricity, traffic, and finance) demonstrate that CM-TFD achieves superior performance compared to state-of-the-art baselines. The results not only verify the model's effectiveness and generalization capability but also highlight its robustness against noise and perturbations.