Rowing load and performance modeling by integrating random forest algorithm

To improve the scientific level of rowing training load regulation and performance prediction, it is necessary to construct a modeling method that can be used to assist in the optimization of individualized training. This study proposes a rowing load and performance prediction framework that integrates multi-source training data and is suitable for individual modeling. First, it integrates self-collected physiological and training data while integrating two public datasets to develop a transfer verification system, ensuring data diversity and result universality. Second, based on a sliding window mechanism, it extracts multi-modal features, including heart rate (HR), stroke rate, training density, Rating of Perceived Exertion (RPE); meanwhile, it encompasses innovative indicators such as the Stress-Recovery Index and performance fluctuation factor. Finally, this study designs a group-individual two-layer random forest modeling framework, integrates linear regression to improve the ability to capture short-term trends, and introduces an individual fine-tuning strategy to enhance prediction accuracy. Experimental results demonstrate that the constructed model’s root mean square error on the test set is 2.94, which is better than that of models like eXtreme Gradient Boosting (3.12) and Long Short-Term Memory (3.67). The average error of the individual model is 27.2% lower than that of the group model. SHapley Additive exPlanations analysis indicates that RPE, HR recovery, and training density are the main predictive factors, with a consistent importance ranking. The study shows that the proposed method has good accuracy and generalization ability in sports performance prediction, and is expected to provide data support for formulating rowing training plans. This modeling strategy has positive significance for individual modeling in sports science and designing intelligent decision-making systems.