We propose an effective time-series prediction method based on decomposed forecasting strategy and channel-independent linear model design. This novel sequence prediction method significantly improves the common prediction distortion issues in small datasets. Meanwhile, it effectively reduces model complexity and runtime using channel-independent linear model design. It consists of two key components: (1) dividing the complete time series into multiple segments to learn time trends and eliminate prediction distortions caused by insufficient data through multiple predictions; (2) adaptively adjusting Patch Size to fit different lengths of subsequences, capturing effective information of different length sequences better with different granularities of Patch. Based on these two components, we propose Multi-scale Time-series Feature Fusion Net (MTFNet), which adopts a three-level processing structure, decomposing the entire prediction process into range partitioning, processing modules, and weighted averaging. The first level finely partitions the input sequence to capture local feature accurately, dividing the sequence into multiple subsequences for local information modeling. The second level processing module preprocesses and linearly transforms each subsequence to further extract and process local features. Finally, the last level filters subsequence results through weighted averaging, obtaining the final prediction by considering the feature information of each subsequence comprehensively. MTFNet achieves competitive predictive performance on 8 real-world datasets, outperforming various state-of-the-art and baseline models on most datasets.