Diabetes is a common chronic disease， and it is very important to control blood sugar for preventing diabetes. However， the uncertainty of continuous glucose monitoring （CGM） data extraction significantly increases the difficulty of blood glucose prediction. Therefore， this article proposes a new deep learning based blood glucose concentration prediction model， aiming at improving the model’s adaptability to sensor extracted data. In this model， the stacked denoising auto encoder （SDAE） is embedded into the structure of the Transformer encoder to achieve reconstruction， denoising， and feature extraction of input data. Then， a mixed position encoding strategy is adopted to replace the original single absolute position encoding embedding， and a lightweight decoder is introduced into the Transformer model to replace the original structurally complex decoder， aggregate feature information from different levels， and obtain local and global features simultaneously. Finally， by constructing an improved SDAE-improved Transformer network for parallel training of CGM data sequences， temporal patterns and complex correlations in the data can be more comprehensively captured， thus improving predictive performance. Experimental results show that the model has achieved significant performance improvement in blood glucose prediction tasks compared to traditional methods， confirming its effectiveness and robustness in processing CGM data.