Modelos de aprendizado de máquina para estimativa de temperatura de operação de módulos fotovoltaicos em clima semiárido

Abstract

The growing demand for clean and renewable energy sources drives the expansion of photovoltaic (PV) systems in semi-arid regions, where high solar irradiance is coupled with high temperatures that reduce energy conversion efficiency. The operating temperature of PV modules (T mod ) is a critical variable for sizing, monitoring, and performance optimization. This work implements and evaluates Machine Learning (ML) models for estimating T mod under tropical semi-arid climate conditions. Using experimental data collected at the Alternative Energy Laboratory of the Federal University of Ceará (LEA-UFC), in Fortaleza, during two years of monitoring (May 2023 to May 2025), models organized into three categories were implemented and compared: ML algorithms (Linear Regression - LR, K-Nearest Neighbors - KNN, Random Forest - RF, and eXtreme Gradient Boosting - XGBoost), hybrid semi-empirical correlations (linear, non-linear, and rational), and consolidated physical-empirical models (Ross, Faiman, and PVsyst). The final dataset comprised 273,310 observations over two years, using five meteorological predictor variables: solar irradiance (G), ambient temperature (T a ), wind speed (V w ), and solar angles. The results demonstrated the superiority of ML models, particularly KNN, which achieved a Mean Absolute Error (MAE) of 0.83°C, Root Mean Square Error (RMSE) of 1.44°C, and a coefficient of determination R 2 = 0.97, representing a reduction of approximately 77% in MAE compared to physical models. The hybrid semi-empirical correlations showed intermediate performance, with a 22% reduction in RMSE and a 68% reduction in systematic bias compared to conventional physical models. Feature importance analysis revealed that G and T a account for 70–80% of the predictive capacity, while V w contributed less than 2%, a consequence of the urban microclimate with high roughness and low average speeds. The superior performance of ML is explained by the stability of the local semi-arid climate, which facilitated pattern learning, and by the algorithms’ ability to automatically adjust to local data, capturing complex thermal variations under high irradiation that generic physical models fail to predict. This work deepens the understanding of PV module heating in the Brazilian semi-arid region, providing a foundation for optimization, cooling, and intelligent control strategies.