Grain storage temperature prediction is crucial for silo safety and can effectively prevent mold and mildew caused by increasing grain temperature and condensation due to decreasing grain temperature. However, current prediction methods lead to information redundancy when capturing temporal and spatial dependencies, which diminishes prediction accuracy. To tackle this issue, this paper introduces an adaptive multi-scale feature fusion transformer model (AMSformer). Firstly, the model utilizes the adaptive channel attention (ACA) mechanism to adjust the weights of different channels according to the input data characteristics and suppress irrelevant or redundant channels. Secondly, AMSformer employs the multi-scale attention mechanism (MSA) to more accurately capture dependencies at different time scales. Finally, the ACA and MSA layers are integrated by a hierarchical encoder (HED) to efficiently utilize adaptive multi-scale information, enhancing prediction accuracy. In this study, actual grain temperature data and six publicly available datasets are used for validation and performance comparison with nine existing models. The results demonstrate that AMSformer outperforms in 36 out of the 58 test cases, highlighting its significant advantages in prediction accuracy and efficiency.