Predictive analysis of porous media–cooled photovoltaic panels using gradient-boosting machine learning models

This study develops a robust machine learning framework to predict the temperature and power output of PV panels cooled with porous media. Four advanced gradient-boosting algorithms, CatBoost, XGBoost, LightGBM, and GBM, were evaluated using five progressively complex models that incorporate key cooling parameters: solar radiation, channel height, coolant type, porosity, flow rate, and ambient conditions. Predictive performance was assessed using multiple metrics, including mean squared error, mean absolute error, coefficient of determination, Pearson correlation coefficient, Nash–Sutcliffe efficiency, Willmott's index of agreement, 95th percentile uncertainty, as well as Taylor diagrams and violin plots to evaluate residual distributions and uncertainty. Results indicate that predictive accuracy improves substantially with the inclusion of additional relevant features. CatBoost demonstrated the highest accuracy and reliability, achieving R2 = 0.95, MSE = 0.45, MAE = 0.32, and the lowest absolute relative average error (0.38 %) against experimental data. XGBoost showed comparable stability with R2 = 0.94, particularly in residual distribution and uncertainty analyses, with both models providing generalized predictions closely centered around zero error. Violin plots and Taylor diagrams confirmed strong agreement between predicted and actual PV output power, with CatBoost achieving the lowest uncertainty bounds (U95 = 1.0075).