Hybrid Transformer-LSTM Model for Athlete Performance Prediction in Sports Training Management

Precise athlete performance prediction is required for the optimization of training regimens, the prevention of injuries, and the improvement of performance in competition. In this study, we propose a self-attention mechanism-based transformer LSTM (HTL) Athlete Performance Forecasting (APF) framework using Transformer along with Long Short Term Memory (LSTM) networks to model sequentially. The framework can capture global feature interactions and localized temporal dependencies in athlete performance data. A dataset containing 200 football, basketball, and athletics athletes over 12 months was used to train and evaluate the model. Heart rate, speed, distance, workload and recovery indicators are performance metrics. Indeed, HTL-APF is validated against baseline models such as Transformer only, LSTM only, CNN, and RNN models at segmenting the sequence via a sliding window approach. Precision, Recall, F1-Score and AUC-ROC are the evaluation metrics. We propose HTL-APF that results in an F1-Score of 92.1%, AUC-ROC of 96.3%, which outperforms the Transformer model (F1: 88.1%, AUC-ROC: 92.4%) and Lstm only model (F1: 85.9%, AUC-ROC: 90.1%). Analysis from class to class reveals that there is a high classification accuracy (97%) for top performers and moderate (89%) and bad (90%) performers also have good performance. In addition, precision and F1 scores for cross-domain testing across sports disciplines remained above 91%, indicating the framework's generalizability. HTL-APF is a scalable and accurate solution for athlete performance forecasting for personal training plans, injury prevention, and real-time decisions in sports training management, illustrated by these results. Given that it is intended for real-world sports analytics, future work will investigate the development of lightweight adaptations, enhanced interpretability, and domain-specific extensions to enlarge its application range.