To address issues in athletic training such as load scheduling relying on experience, difficulty in timely quantifying performance fluctuations, and a disconnect between training recommendations and risk identification, this paper proposes PTO-Net, an intelligent computing-based model for optimizing athlete training.Based on publicly available SoccerMon monitoring data, the study unifies external load, internal load, recovery status, and contextual factors at the athlete-day granularity to construct continuous training samples. On this foundation, it achieves joint modeling of risk warning, readiness prediction, and load adjustment. This model utilizes time-series weighting method for depicting the recent changes of training, brings in individual baseline correction to consider the differences in athletes’ tolerance ranges, hence generates continuous training suggestions through a multi-task output layer. Experiment outcomes prove that PTO-Net attains the best effect on all main tasks: Risk warning AUC reaches 0.871, Risk F1 reaches 0.793, readiness prediction MAE and RMSE are 4.38 and 5.63 separately, load adjustment MAE is 0.087, and directional consistency ratio is 83.2%, hence all of these are exceed the performance of RF, XGBoost, LSTM, Transformer baseline methods. Three-dimensional response surfaces further point out that the overlapping area of low recovery and high variation corresponds to more obvious suggestions for cutting down training quantity. The outcomes of ablation experiments show that the restoration branch, the single baseline revision, and the time weight together are the foundation of the merits this model has. The method which is put forward in this paper gives continuous support for pre-meeting check, training day arrangement, and recovery monitoring, and thus provides a usable calculation framework for the detailed carrying out of wise choice making in sport training.
