In order to ensure that the physical function of football players adapts to the development of modern football level, and avoid the phenomenon of inability to adapt to the intensity of modern football games due to lack of physical fitness. Aiming at the physical training of football players, this paper proposes an improved long-short-term memory network (W-LSTM) model for the optimization and prediction of physical training. The model effectively combines the global feature extraction ability of LSTM for time series data and the preprocessing ability of the extracted data, which reduces the loss of feature information and achieves high prediction accuracy. The front door is added on the basis of LSTM, which combines training and physical function to reduce the impact of fluctuations in data outliers on the prediction results, effectively improving the accuracy of physical training optimization and prediction, and using body shape, exercise tolerance, exercise intensity and fitness level as input values to conduct comparative experiments on the three models of W-LSTM, LM-BP and ARIMA. The study found that W-LSTM has a lower mean square error (0.011) and a higher correlation coefficient (0.985), indicating that the model proposed in this paper is significantly better than other existing comparison models in terms of the accuracy of prediction results.