<p>Time series forecasting remains challenging due to the presence of non-stationary dynamics, multi-scale temporal patterns, and heterogeneous data characteristics. To address these issues, we propose a novel framework that integrates multi-scale temporal decomposition with a Mamba–Attention hybrid architecture. The decomposition module separates the input sequence into trend, seasonal, and residual components to better capture distinct temporal dynamics, while the HybridBlock leverages Mamba to model long-range dependencies and self-attention to capture fine-grained local patterns. Extensive experiments on urban air quality benchmark datasets demonstrate that the proposed approach achieves superior forecasting accuracy and computational efficiency compared with several state-of-the-art methods. Furthermore, ablation studies confirm the effectiveness of the decomposition mechanism and the hybrid architecture, highlighting their critical roles in improving model robustness and overall predictive performance for time-series forecasting tasks.</p>