<p>Coal mine transportation safety is of vital importance. Traditional coal flow prediction methods have problems such as low accuracy and susceptibility to environmental factors. To address these pro blems, this paper proposes a multi-scale residual convolutional autoencoder (MR-CAE) for coal blockage detection in transfer stations. The model introduces a multi-layer residual convolutional block to improve the decoder, combined with a multi-scale enhanced fusion convolutional block (MEFCB), and uses a dynamic attention mechanism to adaptively fuse coal flow features to obtain higher detection accuracy. Finally, skip connections are used to enhance the model's adaptability to the environment. Indicators such as MAE and MSE are used to evaluate the prediction performance, and the robustness of the model is ensured by testing in more scenarios. Compared with Swin Transformer and ViT, the prediction accuracy of the model is improved by 3.6% and 3.3%, respectively, and the detection success rate reaches 93.5%. This demonstrates its advanced capabilities in coal flow prediction and provides a reliable framework for coal flow monitoring in transfer stations.</p>