Physics‐Informed Ensemble Learning for Accurate Performance Prediction of Uncharacterized Flat‐Plate Solar Collectors

ABSTRACT Accurate monitoring of solar thermal systems remains challenging in data‐scarce environments, particularly in developing regions where technical documentation, flow‐rate measurements, and reliable sensing infrastructure are often limited or unavailable. Over years of operation, many installations also lose essential manufacturer information, rendering conventional model‐based approaches impractical. Classical thermal models rely on collector‐specific parameters such as η 0 , a 1 , and Ac, while also assuming steady operating conditions that rarely persist under real outdoor dynamics. To address these limitations, this study introduces a manufacturer‐parameter‐independent predictive framework based on physics‐informed ensemble learning. A dataset of 13,948 experimental observations collected from a solar thermal platform in Nouakchott, Mauritania, was used to train and evaluate six ensemble learning algorithms under dynamically varying operating conditions. The proposed models predict three key performance indicators: outlet temperature, temperature lift, and normalized thermal efficiency. The feature space combines only field‐accessible measurements—solar irradiance, ambient temperature, and inlet temperature—with thermodynamically meaningful variables derived from the EN 12975 reduced‐temperature formalism, including the reduced temperature T*. Among all evaluated approaches, the Gradient Boosting Regressor (GBR) achieved the best overall performance for outlet‐temperature prediction, reaching R 2  = 0.9635, MAE = 0.61°C, and RMSE = 0.83°C under five‐fold cross‐validation. Comparable accuracy was obtained for temperature lift ( R 2  = 0.91, MAE = 0.61°C), while the normalized efficiency proxy was predicted with near‐perfect accuracy ( R 2  = 0.997, MAE = 0.001). Physics‐based feature augmentation produced a statistically significant improvement in predictive capability compared with raw sensor inputs alone. SHAP‐based interpretability analysis further revealed that solar irradiance and the reduced temperature act as dominant explanatory variables, indicating that the models implicitly capture key thermodynamic and thermal‐inertia behaviors. Temporal autocorrelation analysis further confirmed that T* retains statistically significant short‐term thermal memory, with a characteristic persistence time of approximately 85 min, and this memory propagates into consistently positive lagged cross‐correlations with outlet temperature. These results demonstrate that physics‐informed ensemble learning can provide accurate, robust, and interpretable monitoring of flat‐plate solar collectors without requiring manufacturer‐specific parameters or extensive instrumentation. The proposed framework therefore offers a scalable pathway toward virtual sensing, fault‐aware monitoring, and digital‐twin development for solar thermal systems operating under real‐world conditions.