Existing time series prediction methods typically use entire time series sequences as input for prediction, neglecting the importance of multi-granularity understanding of time series. This paper proposes a Multi-scale Fusion Recurrent Attention prediction model (MFRA) with three key innovations: First, it designs an encoder that integrates convolution and attention mechanisms to capture both local and global dependencies in the input sequence, improving prediction accuracy. Second, it introduces a Bottom-Up Multi-scale Temporal Feature Pyramid architecture that hierarchically decouples time series while combining convolution and attention mechanisms, enabling the model to simultaneously focus on short-term fluctuations, medium-term trends, and long-term periodic patterns, enhancing prediction efficiency and accuracy through a bottom-up fusion mechanism. Finally, it designs an End-to-First attention module that captures rich periodic patterns embedded in multi-scale fusion features by dividing periods and constructing cyclic period sequences. Experiments on multiple benchmark datasets demonstrate that MFRA achieves significant performance improvements compared to baseline models, particularly evident in long sequence prediction tasks with larger data volumes. Our code and model are released in https://github.com/manxswl/MFRA/tree/main .