Time series forecasting plays a very important role in a variety of task-specific applications. However, real-world time series data typically show different patterns on different cycles, and changes in trends at future moments are influenced by the interaction of these overlapping multi-period data. The existence of stable periodic patterns in time series data is the basis for long series forecasting. Based on the multi-periodic patterns in time series, this paper proposes a new adaptive multi-period time series analysis model MAPformer. In order to fully exploit the inherent multi-periodic patterns in time series data, the fixed periodic patterns of the data are eliminated before entering the Residual Aggregation Former (RAF), so that the RAF focuses on the local trend of the data within the period, and finally the periodic patterns of the data are added back in the feature aggregation stage. In order to extract both long-term inter-periodic variation and short-term intra-periodic variation of the time series data, the model uses Multi-Periodic Transpose Block (MTB) to reconstruct the time series data into a 2D sequence, so as to extract the periodic features of the time series data from multiple dimensions. Finally, Periodicity Mixing Machine (PMM) is used to fuse short-term intra-periodic trends with long-term inter-periodic trends in order to explore the cross-periodic modeling process. Comprehensive experiments demonstrate that MAPformer not only overcomes the limitations of existing methods, but also achieves optimal performance on most datasets.