In recent years, channel-independent linear models have demonstrated strong performance in multivariate time series forecasting, thus emerging as a promising solution for lightweight modeling due to their computational efficiency and robustness under limited data. However, by neglecting inter-variable dependencies, they remain constrained in their ability to capture complex cross-channel dynamics. To address this limitation, we propose DAF-Net, a novel multivariate time series forecasting framework that integrates both a Spectral-Dynamic Aggregation Redistribution (SDAR) module and a Cycle-Aware Frequency Mechanism (CAFM). The SDAR module enhances cross-channel interactions by combining quantile-based spectral filtering with a learnable sparse channel weighting matrix, while CAFM dynamically amplifies salient frequency components through phase modulation and frequency gating, and further leverages period-aware embeddings and mirrored spectral reconstruction to robustly capture periodic structures and frequency drift in time series data. To improve robustness against distribution shifts, DAF-Net incorporates Reversible Instance Normalization (RevIN) to stabilize training and ensure distributional consistency. Extensive experiments on diverse real-world benchmarks demonstrate that DAF-Net consistently outperforms state-of-the-art forecasting models in terms of both accuracy and generalization.