This study tackles the challenges faced by current physical education training methods in real-time monitoring, data transmission, and intelligent analysis. We introduce a numerical simulation and optimization approach for physical education training, leveraging embedded wireless communication networks. By combining Atrous Spatial Pyramid Pooling (ASPP) and Long Short-Term Memory (LSTM) networks, our method effectively processes and analyzes athletes’ multi-scale spatial features and temporal sequence data. Experiments on the PAMAP2 and MHEALTH datasets show that our approach surpasses other mainstream methods in key metrics such as maximum F-measure, Mean Absolute Error (MAE), weighted F-measure, and structure similarity measure, with a notable advantage in enhanced alignment measure. Ablation studies further validate the contributions of the ASPP and LSTM modules. This method enhances the accuracy and real-time prediction of training outcomes, offering valuable insights for the advancement of intelligent physical education training systems.