This study develops data-driven DC power prediction models based on high-frequency measurements from a demonstration OFPV project in the Yellow Sea, China, using three tree-based regression algorithms: Random Forest (RF), XGBoost, and LightGBM. A two-stage hyperparameter tuning strategy—Bayesian optimization followed by grid search—was implemented alongside a customized composite scoring function that integratesR2, RMSE, and MAE to comprehensively evaluate model performance and generalization capability. Results indicate that the RF model achieved the best performance on the test set, with a coefficient of determination (R2) of 0.946, a root mean square error (RMSE) of 302.56, and a mean absolute error (MAE) of 180.11—substantially outperforming XGBoost and LightGBM. Feature importance analysis reveals that frontal irradiance contributes over 65% to power output, followed by sea surface albedo (12%–26%), while module backplane temperature and wind speed have relatively minor effects. This study provides an efficient machine learning approach for OFPV power prediction, offering theoretical support for future engineering applications and model refinement.
Xu et al. (Sun,) studied this question.