Development and validation of interpretable machine learning models for photovoltaic panel temperature prediction

Abstract

Accurate prediction of photovoltaic (PV) panel temperature is critical for optimizing the design, operation, and maintenance of PV systems. Although many steady-state and machine learning (ML) models have been proposed to characterize the relationship between meteorological elements and panel temperature, achieving a balance between prediction accuracy, interpretability, and extrapolation capability remains a challenge. Therefore, this study attempts to construct a PV panel temperature prediction framework that integrates feature engineering and interpretable ML techniques. A feature selection method combining Pearson correlation coefficient, Shapley additive explanations, and extreme gradient boosting quantitatively evaluates the correlation of meteorological elements and their contributions to temperature prediction. Furthermore, two symbolic regression methods based on genetic programming and multi-population evolutionary algorithms are employed to develop explicit models with concise expressions and excellent performance. Two experimental datasets are from a utility-scale PV plant and a commercial rooftop PV system, with sizes of 19383 × 6 and 4503 × 6, respectively. Experimental results show that the proposed method can accurately and reliably predict the operating temperature of different panels, achieving R² of 0.981 and 0.961. Comparative analyses highlight the superior accuracy, interpretability, and broad applicability of the proposed models. This work provides valuable insights for panel temperature prediction, interpretable ML model development, and PV system management.