<p>Time-series forecasting remains a critical task in domains such as energy, power systems, and finance. Although MLP-based models provide lightweight and efficient alternatives to recurrent and attention-based networks, they struggle with capturing temporal dependencies and multiscale dynamics. To address these issues, this paper proposes a novel time-series forecasting framework based on Adaptive Attention Fusion and Multiscale Decomposition (AFMD), built upon a Multi-Layer Perceptron (MLP) architecture. This model innovatively integrates multiscale signal decomposition techniques with a dual-dependency attention mechanism, creating a hierarchical paradigm for temporal feature learning. First, a multi-resolution downsampling strategy is employed to decompose the original time series into subsequences with distinct dynamic characteristics. Next, a parameter-shared sparse gated multi-head attention mechanism is introduced at the subspace level to enable fine-grained modeling of cross-scale temporal patterns. Building on this, a multimodal feature fusion mechanism and dynamic weighted predictor are designed to achieve collaborative optimization of multi-granularity features and probabilistic integration of forecast results through an adaptive gating network. Extensive evaluation on eight real-world datasets demonstrates that the proposed model exhibits outstanding efficiency across various performance metrics, and ablation experiments further validate the effectiveness of the proposed modules.</p>