Time series prediction has a wide range of applications in many fields. In real-world scenarios, time series in different fields show diverse temporal variations and complex features, posing challenges for prediction. Current mainstream prediction algorithms have made some progress based on CNN or Transformer. CNN can effectively capture the local features of the data and the latter can handle the long-term dependencies in the sequence. This paper proposes a new multi-scale framework that integrates CNN and Transformer. According to a set of period lengths, we converted the 1D time series into 2D tensors, with intra-periodic and inter-periodic variations corresponding to the rows and columns of the 2D tensor, and sampled by 2D kernels step by step, obtained multiple sets of 2D tensors with different views. The sampling depth corresponds to the macroscopic and microscopic variations of the data and then models the long-term dependencies with the transformer. Then fusing the multiple sets of features at different scales to obtain the final prediction. Thus, our approach combines the advantages of CNN and Transformer for time series forecasting and outperforms other SOTA works in both long-term and short-term prediction.